💬｜general-chat

[buoyant moss]

@shell tendon I use diffusers with DPM++2M

Around 7 to 8 seconds for 50 steps on SDXL base

[vagrant fox]

14 of what? iterations/s or what

[fervent thunder]

yes

7680 cuda cores

[buoyant moss]

At 1024x1024 res

[fervent thunder]

So either they are lying or I am missing out on a lot of performance

No no at 512 512

[shell tendon]

gotcha, so about 6-7 it/s

[buoyant moss]

No, I mean for me 50 steps of SDXL base + DPM++2M + 1024x1024 is 7~8 seconds on a 4090

[fervent thunder]

ah

[charred mesa]

that was supposed to be the discord server, but as jayjay is saying, today might also be a day for that

[vagrant fox]

my 4080 super got 6-8 it/s on juggernautXL at 1/4th resolutiuon of 4k

on auto1111

[fervent thunder]

on XL 1024x1024 i am getting like

under 2 it/s

[buoyant moss]

Ooof that's really really bad

[fervent thunder]

i thought that was normal until i read about other ppl's perf

[buoyant moss]

Doesn't sound right

[vagrant fox]

u run command no half or some shit in your bat file?

[fervent thunder]

nothing

just default

[buoyant moss]

That sounds like he's running at fp32

Instead of the faster fp16

[fervent thunder]

a1111

[charred mesa]

ohh

[fervent thunder]

how do i switch

to that

[shell tendon]

just checked, dpmpp 2m on forge i get about 6.5-7 it/s

[buoyant moss]

What gpu?

[shell tendon]

4090 gigabyte oc

[buoyant moss]

Nice

That sounds about right

SD3 8B being 5x slower is going to hurt thou 🙂

[vagrant fox]

not really if you do just images, imagine hurt while making 9k frames of animation

but still you dont need t5 for every generation

[buoyant moss]

I am curious why they used T5, I'd probably have used a multi-lingual model like BGE-M3. Although the M3 just came out last month

https://huggingface.co/BAAI/bge-m3

[vagrant fox]

most likely because they work on sd3 longer than a month, trained this shit for long time

[buoyant moss]

Yeah fair

[pine fiber]

isnt it a lot smaller?

[buoyant moss]

Embedding models are typically smaller than auto-regressive text models

https://miro.medium.com/v2/resize:fit:1400/format:webp/1*x_IbdvfVCSm1OLaP3Ev72w.jpeg

https://towardsdatascience.com/openai-vs-open-source-multilingual-embedding-models-e5ccb7c90f05
Third party benchmarks show that BGE-M3 is better performing than even closed source OpenAI embedding models

[shell tendon]

i have a few questions regarding that

are they defining the generation time in the way people often do with sd15 or sdxl... 20 steps with the fastest sampler?

that's q #1

[buoyant moss]

The times they give are 50 steps on 4090

That's all we know

[shell tendon]

q #2 is something i'll have to answer myself... is the improvement in image quality and prompt adherence sufficient to exceed the value of a batch size of 4?

[vagrant fox]

1024x1024 50 steps, no sampler mentioned

[shell tendon]

ahh k

cuz if it does, then we're right back to the same run times more or less

[buoyant moss]

Stability claims 34 seconds for 50 steps on 4090 for SD3 8B

[shell tendon]

my guess is it'll save time because of that

[fervent thunder]

oh shit

I slept on a driver update

Could that really be it though

i doubt it

[buoyant moss]

You're using windows?

[fervent thunder]

i am

[buoyant moss]

Wow

[fervent thunder]

i dont think thats it, i have the one from 22 feb

[buoyant moss]

I'm genuinely surprised. Not being sarcastic or anything

[vagrant fox]

yeah some of us like to play games so linux is not option

[fervent thunder]

why are u surprised

and what does that have to do with my low it/s

[buoyant moss]

... I dual booth Windows + Linux

[shell tendon]

i'm on windows myself but that's because i have so much work stuff i need throughout the day

[pine fiber]

not everyone has a degree in cs just to use linux ;(

[shell tendon]

where the software is windows only

no need for a cs degree to use linux

[pine fiber]

I tried installing linux once and it bricked my entire pc had to factory reset

[shell tendon]

it's pretty easy these days tbh

wtf

i'm waiting for the next amd chip to come out

[vagrant fox]

GL running most of games on linux, LTT tried that not that long ago and experience suck harder than sasha grey

[shell tendon]

then i'm building another system, moving my 4090 over there, putting the 3080 12gb back in this one (it's just sitting on the shelf collecting dust atm)

putting linux on the 4090 and leaving windows on this one

[pine fiber]

u think 5000 series will be any good?

[shell tendon]

i haven't run SD on linux but i've heard it's great

[fervent thunder]

what does using --xformers do?

[vagrant fox]

5090 rumoured to be 70% faster than 4090

[pine fiber]

or does nvidia not really care about consumer line anymore

[fervent thunder]

i dont see any difference with it

[shell tendon]

i've heard all kinds of conflicting things about the 5090

[buoyant moss]

it uses xformers

[shell tendon]

that it'll be 24gb, 32gb, or 48gb

[fervent thunder]

48gb i doubt

[buoyant moss]

Optimized CUDA kernels for self attention

[fervent thunder]

that is a huge leap

[shell tendon]

that it'll be an incremental increase in speed, and that it'll be almost twice as fast

or even more

[vagrant fox]

if they give only 24 gb on 5090 then fuck nvidia

[fervent thunder]

I am more expecting same vram but much faster memory

but 32gb is possible too

[shell tendon]

one rumor was 50k cuda cores (4090 has 16k or so, 4080 10k or so, etc.)

[vagrant fox]

its new samsung memory on 5090

[shell tendon]

yeah

[fervent thunder]

50k would be

insanity

in price

[shell tendon]

if the ram bumps to 32gb or higher i'm going to get one immediately

[fervent thunder]

im already saving up

for the 5090

[vagrant fox]

it won't be 50k cuda cores nvidia wont shoot themselves in leg, if 5090 has anything like 25-30k cuda cores its upgrade worth getting still

[fervent thunder]

or 5080 if 90 costs like $3000

[vagrant fox]

90 will be 2-2.5k usd for sure

[fervent thunder]

price is decided last i think

i can just hope

[rich kestrel]

wont be that expensive

[fervent thunder]

cause in my country you pay way above msrp due to the insanely high vat

[gleaming spire]

Can someone assist me here my images keep coming out really pixelated for some reason after installing SDXL

[vagrant fox]

it will be 2k at reelease i bet esily considering how prices are hiking

[fervent thunder]

show the image and ur settings

[rich kestrel]

change the vae

[shell tendon]

that sucks...

[vagrant fox]

they wont give you new type of memory and boost cuda cores 50% for same price mark my words

[fervent thunder]

like we are talking 55-60% vat

[shell tendon]

wow.

[fervent thunder]

i paid over 4090 MSRP for my 4070 ti xD

4090 was over 3000 eur

[shell tendon]

the 50k would be more than tripling if true

[fervent thunder]

4080 2500

[vagrant fox]

where u live north korea?

[gleaming spire]

check your dms I sent SS @fervent thunder

[fervent thunder]

serbia

[shell tendon]

yeah that's crazy

[fervent thunder]

k

[vagrant fox]

it wont be true they wont give u so many cuda cores they still want to sell proffesional grade gpus

[shell tendon]

yeah they need real competition

[vagrant fox]

thats why they always limit vram and cuda cores

[shell tendon]

they would if they had it

but they don't right now so your perspective is the one i share

i wouldn't be at all surprised if they drop another fn 24gb on us, 48gb is a total no

[vagrant fox]

no, you wont get 50k cuda cores gpu for 2k usd in 5000 series even if amd or intel are more competetive

[shell tendon]

best case scenario is 32gb with maybe 20k cuda cores and faster vram imo

[fervent thunder]

ye

that sounds realistic

or 20-somethingk

in the lower 20s

[vagrant fox]

it will have more than plain 20k, 2090 had 4352 CUDA cores, 3090 had 10496, 4090 had 16,384

[fervent thunder]

they promised the biggest generational leap yet

implying bigger than 4090 vs 3090

[vagrant fox]

i would expect something like 22-27k

[fervent thunder]

if that is true

it is going to be one hell of a card

[vagrant fox]

yes there is rumour that it will be 70% faster than 4000 series card

[fervent thunder]

it jsut remains to be seen

what the price difference will be lol

[vagrant fox]

+500 usd to 4090s msrp for sure

for 5090

esily

[fervent thunder]

that is what the trend has been

roughly 33% price increase

for a massive perf increase

[vagrant fox]

also GDDR7 memory, that will be costing them for cutting edge new type of memory

[fervent thunder]

Tbh i might just go 5080 if it follows the same increase from 4080

Its more than enough for me

for 165fps @3440x1440p

[vagrant fox]

thats why i got 4080super month ago instead 90 for now, but lets see how it goes if i generate through ai enough income i might jump onto 5090, got atx 3.0 1500w psu so i am ready

well for gaming at 2k my 3080 was enough

[fervent thunder]

damn 1500w

[vagrant fox]

im not upograding cards for gaming anymore just for ai

[pine fiber]

shits gonna cost $4000

[fervent thunder]

bro have u tried a qd oled monitor

if ur investing into top tier hardware

u gotta try the top tier monitors

im never going back to lcd

[vagrant fox]

im on samsung odyssey g7 27 inch 2k 240 hz

[fervent thunder]

I recommend this one

https://www.dell.com/en-us/shop/monitors-monitor-accessories/ar/4009/qd-oled?appliedRefinements=40622

[vagrant fox]

alienware have top tier monitors now for panoramic view

[fervent thunder]

i think there is a 240hz model as well but i could be wrong

there is also 21:9 one

that i have

but 165hz

the screen looks amazing though

it is an even bigger wow factor than going from 60 to 240hz

[pine fiber]

i have a 360hz qd oled

its alright

wouldnt recommend it

[fervent thunder]

that is an understatement

what do u recommend then

[vagrant fox]

i think leap from 240 to 360 is less noticable than from 144 to 240

[fervent thunder]

240 to 360 is small

diminished returns at that point

[vagrant fox]

also i am getting older im not playing counter strike like before so dont need that kind of refreshrate, ill mark 240 as my standard to live on

[fervent thunder]

360 wont make anyone a better player than 240

[vagrant fox]

i think ill swap to 4k monitor when cards can esily hold 144 ultra settings in most aaa games, for now on im sticking to 2k for few more years

[fervent thunder]

the other features about these 360hz monitors are more appealing

like the black frame insertion

or qd oled tech itself

there is a 4k 240hz qd oled already i believe

or one coming soon

[vagrant fox]

well top tier players will feel difference LTT took shroud to test it

[fervent thunder]

be on the lookout

u can feel the difference yes

[vagrant fox]

but how many players are like shroud, 1 in 50 million?

[fervent thunder]

i tested it myself

it is very subtle though

its nothing like 144 vs 240

240 makes 144 feel slow as fuck

but 240 still feels ok compared to 360, 360 is jsut a lot smoother

[shell tendon]

man, i have a neo g9 (max refresh 240) and i cant' really tell

but i also don't play competitive shooters

[fervent thunder]

but if ur buying one of those zowie 360hz monitors

the bigger difference will be the BFI tech

[shell tendon]

once it hits like 40-50 fps i'm usually happy enough

[vagrant fox]

well i feelt ddifference 144 vs 240 in counter strike but it was at times i did play shit ton of it at not bad level with doing a lot of flicks (quickscope kills with awp)

[shell tendon]

anything below 40 is too low for me though for anything

ahh

[fervent thunder]

I mean u dont need to play competitive games to tell the difference

But for singleplayer games 130-140fps is the highest i'd take

before i start upping settings

[vagrant fox]

depends on game i can play RPG with 60 fps, but racing games etc i need at least 100+ to feel game properly

[fervent thunder]

like cyberpunk, alan wake 2 ill happily play at 60 fps with maxed everything

[vagrant fox]

as long as fps dont suck like on switch or consoles overall lol

zelda with drops to 18 fps lol

[shell tendon]

lol yeah pretty bad on the switch

40 vs 30 is huge

[fervent thunder]

i mean it is switch

[vagrant fox]

ps4/xbox sucked a lot too

[fervent thunder]

visually 30 fps is fine but when you are the one playing it

the input lag is really horrible

[vagrant fox]

ps5 is bit better finally got rid of HDD and got frame generation and upscaling but still sucks big time to be on that hardware

[fervent thunder]

80+ fps is where it starts to feel normal for me

[shell tendon]

that's 33 ms between frames vs 25... 50 fps = 20 ms between frames

[fervent thunder]

like i dont have input lag

[shell tendon]

that's an important line to me cuz of having done a bunch of audio work in the past

[astral goblet]

ever since i started with the vive vr headset, i get motion sickness when my frames don't match my screens refresh rate. i hate it so much. 100 fov (depending on the engine) helps me a lot. I played games for 30 years without motion sickness and ever since VR started triggering it for me, regular gaming does too! hates it so much

gotta eat ginger and friggin wear c-bands to get a good fps session in

[fervent thunder]

i dont have a headset but i turned on gsync a while ago

didnt feel much of a difference

[shell tendon]

if you put a sound on a delay line and the echo is around that 20ms or less point, it sounsd like chorus, not an echo

[fervent thunder]

until one day i turned it off

holy shit was it disgusting

immediately turned it back on

[shell tendon]

my brain at least seems to do something similar with images

[vagrant fox]

fuck vr im waiting for proper ar glasses with small form factor so i can have virtual monitors floating around my room while still using my main monitor for low imput lag etc, something like apple vision pro but refined

[astral goblet]

lol once you're immersed in that beautiful smooth refresh, going back is rough

[fervent thunder]

especially on oled with the low response

[vagrant fox]

vr is immersive but i had cold sweat after 1 round in fake counter strike when i used stick to mvoe around, i had to play games stationary or teleporting

[fervent thunder]

just panning the camera with high fps and gsycn on

feels so nice

[astral goblet]

i mean sure, but thats more AR.

VR is cool in a few cases. I sure havnen't bought a newer headset since the vive though

[fervent thunder]

i tried VR in a gaming cafe

it was dope

it really feels like u are there

[vagrant fox]

yeah ar is for me more usable than vr with it's motion sickness

[astral goblet]

Flight sims , elite dangerous, oh man. so good. Being able to look around in the cockpit to pin your target, whiel using other controls to fly that way, so fun

[fervent thunder]

i bet racing games are awesome

i played a horror game

[astral goblet]

racing games in vr make me puke almsot instantly. since eveyrthing moves but you don't feel it

[fervent thunder]

i think outlast 1

yea i can imagine

it ws trippy to me at first too

[astral goblet]

i press gas and go 5m and i gotta decouple and run to the toilet

[fervent thunder]

i had to adjust

[astral goblet]

i can't

[vagrant fox]

lol

[astral goblet]

i dont know what it is about racing games. flight sims are fine. same idea aren't they?

[fervent thunder]

no

clearly not lol

[astral goblet]

must be the peripheral detail

[vagrant fox]

in flight sim u dont have many obcjects close to you that you are passing

you just surf mostly in sky or look at city from 4000 meters

they would have to pay me to drive in vr

[astral goblet]

i do love the seperation of where you're looking and where you're aiming. i figure with the vive, part of it is screen resolution

[shell tendon]

what i've always wanted is a VR racing booth that can actually convince me i'm driving my usual route to work or the grocery store

i've done these routes so many times i've got the gear changes timed up perfectly on every turn, and have most of the traffic lights memorized...

i'd love to be able to do shit like head for a traffic light at 70mph without slowing cuz i'm confident i timed the change to green perfectly

knowing that if i didn't... just reload unlike life

[vagrant fox]

whats the point of playing game to do same stuff u do irl

[shell tendon]

^^"actions without consequences day"

[vagrant fox]

i would do cannon ball race if i had stamina to handle vr

[fervent thunder]

The irresistible urge to dash into the street, toggle sandevistan, and shoot every driver in the head

and then reload the save like nothing happened

[vagrant fox]

drive over every pedestrian

thats why i love gta

[shell tendon]

yup lol

[fervent thunder]

yep

[vagrant fox]

so terapeutic

[shell tendon]

but really, i'd love to see just how fast i could actually get to work

[vagrant fox]

and they say games create killers

[shell tendon]

if i'm not afraid of a reckless driving arrest

[vagrant fox]

if not games i would most likely be serial killer by now

[fervent thunder]

cyberpunk is my gta at this point

i just like the first person view and graphics

[shell tendon]

timing up blasting through red lights based on a quick glance of cross traffic in both directions

shortcuts through parks, sidewalks

[vagrant fox]

i need to check how it goes year ago there were still plenty of bugs i could not handle i pushed game for story was so good

[fervent thunder]

cyberpunk is in a really good state rn

plus, phantom liberty is as long as the base game story

if not longer

they revamped the whole class trees

[shell tendon]

last time i fired it up i felt instantly bored again

[fervent thunder]

its a completely different game from before 2.0 update

[vagrant fox]

but does npc's still teleport and act like they have palpitation attack or finally they are not? in some moments it was like horro game, npcs teleporting in your face lol

[shell tendon]

feels like a job like so many open world games

checklists, etc

[fervent thunder]

I have not seen that happen

They just take long to notice me

But I do have a stealth build

with like 70% slower detection rate

[shell tendon]

do the cops still appear and disappear like ghosts?

[fervent thunder]

no

cops system is fixed

u got car chases and shit now

road blocks

maxtac

[vagrant fox]

i just go and blast everything or hack them through cameras

[fervent thunder]

which are like a bossfight

it's dope

[vagrant fox]

i will give it a go but they can forget me buying 24 gbp dlc for shitshow they did. waiting for -50% for sure

[fervent thunder]

the only thing they suck at is finding you if you hide

i mean phantom lib is really good

it's longer than the main story

and can even act as an alternative to it

[vagrant fox]

main story was not as long, comparing main story + side quests and making it as equally long would be nice

like elden ring dlc

[fervent thunder]

if u compare it to all side quests + main

its not comparable

there are too many side quests

[vagrant fox]

c'mon there can never be too many side quewsts 😄

btw Outer Worlds was nice rpg i played lately, for sure waiting for second part

[shell tendon]

yeah i used to like that stuff more espec during the lockdowns a few years ago

now i just don't have time

[fervent thunder]

sorry if everyone is ask this, but any word when we can test sd3? i assume it is here with some bots or will be

[shell tendon]

nope

[vagrant fox]

no official date

[fervent thunder]

time for patience something we ai people dont have haha

[shell tendon]

we've already been waiting what, 2 weeks? that's at least 6 months in AI years

[charred mesa]

exactly

[astral goblet]

2 days ago Emad said tommorrow or the next day. That seems official. He's kind of #1 in charge dude

[brittle maple]

hello

[charred mesa]

Emad:

Discord access to do final tunes of model, API access to stress test, model weights drop. May put out inference code shortly too, tidying.
Working on a plan to have the various controlnets etc done

on the subreddit 2 hours ago

[brittle maple]

Can I make a sequential generation in a111 with different promts? 2 promts at a time

[trim nymph]

link to the subreddit please

[charred mesa]

https://www.reddit.com/r/StableDiffusion/comments/1b91fly/comment/ktxuamu/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button

here's the comment itself under the subreddit post

[pine fiber]

he keeps giving random dates and not meeting them, he should stop doing that XD

[charred mesa]

it's been 84 years

[buoyant moss]

You can easily use it as a bash script, and rename the file after you watermark it.

Then in your bash script, check if the file is renamed, and add watermark and rename if not

[shadow herald]

hi

i am new to discord

hi. anyone there ?

[brittle maple]

hi

[shadow herald]

hi dicordos.

pleased to meet you

[vagrant fox]

Hello 🙂

[astral goblet]

no you're a dicordo!

[shadow herald]

Can anyone teach me to how to create an instagram post on discord ?

[fervent thunder]

i dont think that's a thing

[astral goblet]

ones owned by facebook, ones owned by discord. entirely different companies

[fervent thunder]

you would create those on instagram

[vagrant fox]

just create ai image with square aspect ratio

[fervent thunder]

is SDXL supposed to be slower than 1.5?

Even at same resolution

[tough lion]

yes, because it has more parameters

[vagrant fox]

considering xl has most likely more parameters to go through

it would be logical

[fervent thunder]

i see

[astral goblet]

sdxl is supposed to be 1024x resolutions. hard to get quality if you're not using the base attention it was trained for

[fervent thunder]

Cause 1.5 is way faster at same resolutions

So I was curious why

[vagrant fox]

it like running gta vice city vs gta v

gta v would have less fps

[fervent thunder]

yea

[astral goblet]

sdxl fills more memory too. You might just be blowing out your vram and thats why its "Waaay" faster. If you're not using system memory, it should be pretty comparable if not a little slower

how much vram do you have

[fervent thunder]

12gb

[astral goblet]

so a 30 series card?

[fervent thunder]

4070 ti

[astral goblet]

shoudlnt' really be "waaaay" faster for sd15.

[fervent thunder]

oh nevermind

I just realized I kept the same upscale multiplier

but changed res from 512x512 to 1024x1024

far from same res lol

[vagrant fox]

page not found

[astral goblet]

what happened to the same res tests being faster?

\

[fervent thunder]

I did 512x512 upscaled 2x

and then did 1024x1024x upscaled 2x

forgot to turn off the upscale

[astral goblet]

I feel like this is how most help desk troubleshooting goes. Never believe the user. Users always lie

[vagrant fox]

or they don;t know they lie lol

[fervent thunder]

not lying just overlooked the setting

[astral goblet]

forgive them father for they know not what they do

[fervent thunder]

different things flow

[tough lion]

it's just the number of parameters, the sdxl 1.0 has almost 10b parameters https://stability.ai/news/stable-diffusion-sdxl-1-announcement

[vagrant fox]

thats what he meant

[astral goblet]

https://www.tiktok.com/@professortravs/video/7266046841362435370

[tough lion]

while the stable diffusion 1.5 has only ~1B parameters I think, so it's ~10x smaller

[astral goblet]

600m

[charred mesa]

boasting a 3.5B parameter base model and a 6.6B parameter model

[shadow herald]

Can anyone go to the following page on Instagram: https://www.instagram.com/chatgptricks/ and see the page. They have 1.1 million followers. Can anyone tell me what are the secrets to their success ?

[vagrant fox]

follow bots considering amount of engagement

like 98% of big acc's

20k likes with 1.1m followers? yikes

half of likes most likely bought as well

[shadow herald]

what are follow bots ? and how can you buy likes ?[

[tough lion]

yes, this sums up to ~10B

[vagrant fox]

there is plenty of services that offer that, most shit in web is fake and have pumped numbers, viewbots on twitch, follow bots on instagram, fake garbage everywhere

no wonder most of traffic in web is bots

[charred mesa]

you guys use the refiner?

[astral goblet]

bots

[shadow herald]

do you know where i can acquire the service of follow bots ?

[trim nymph]

these are messages for the off topic channel

[astral goblet]

its against terms of service. Tha'ts all dark web stuff. I won't help people be shady

[vagrant fox]

google does not hurt but i dont recommend unless u aiming to resell your page to someone clueless

[trim nymph]

please move your instagram talk to off topic

[astral goblet]

Not a lot of other topics going on in general-chat this morning. I dont think its a big harm

[vagrant fox]

+1

[astral goblet]

please don't minimod

[charred mesa]

lol

[placid wasp]

`Its the evening here. And i like the chat. bots bots bot everywhere

[vagrant fox]

speciallyt when you jump out of blue without engaging with us for last few hours

we won;t, google will

[astral goblet]

never said i would. keep up

take your minimodding elsewhere please. you're not actually a mod

[placid wasp]

google, Ai. the local library (maybe) would all have info

[tough lion]

actually that's why I came here too, the original paper mentions the base + refiner, so you need to do txt2img and then img2img, but many finetunes seem to completely ignore the refiner, so....do most people just use the base and not the refiner?
Even the https://huggingface.co/ByteDance/SDXL-Lightning just does not mention the refiner at all. Same goes for https://civitai.com/models/112902?modelVersionId=126688 do they just ignore the refiner?

[placid wasp]

peeps with Nvida 4090's what are your its per sec onn SD generation locally?

[vagrant fox]

i use gpt for prompt generation for animations

chatgpt

[charred mesa]

Yeah almost nobody uses the refiner

I used it for like the first 1-2 weeks

then just turned to finetuned models

it just takes up way too much ram too

[astral goblet]

why use chatgpt to make prompts? what benefit would that have over "one button prompts"

[vagrant fox]

well i feed small example like 4 prompts (giving full prompt shape with all stuff) then ask for specific animation with prompts every x frames

[vagrant fox]

well try to make 300 line prompt to keep cohherent animation, im too lazy

im not saying to use it for single image

[astral goblet]

i find it blows out community refined models. the details it "refines" are no more or less different from what the base model creates. If you give it more than 5 steps it makes the images haggard

I think the refiner architecture experiment failed. SDXL is just one model. The refiner was just a sunk cost thing they wanted to make work

[tough lion]

haggard?

[astral goblet]

old military slang. The boys in the fox holes look haggard

[placid wasp]

we wern't haggard, we were bedraggled 😛

[tough lion]

oh, I see, thanks.
Hard to say why they did not mention it in the paper, but I guess they used refiner to just queeze better images for the paper? and the improvement of using refiner does not outperform the community finetunes, do I understand it right?

[placid wasp]

or to use a better turn of phrase minging

[astral goblet]

you do your metaphors and i'll do mine.

[astral goblet]

i dont know why they released it. I know it was the first verison of sdxl before they changed to a double clip architectuer, and then they kept it around anyways.

I think base model even works better without it if you use loras

[tough lion]

double clip architecture? I must have completely missed that one

[astral goblet]

sdxl has two clip layers. teh one from sd15 and the one from sd2

[tough lion]

oh, right, I see, I completely forgot it's in the paper

[astral goblet]

if i understand things right, they're basing a lot of their choices on human preference ratings. But that's honeslty just a popularity contest not objective research. What the refiner provides, a lora easily could too. Much lighter weight

[tough lion]

yes, that's true.
Now, when I'm looking at the chart in the paper (oh, I can't paste images here), they have 36.93% wins with just the base, 48.44% winds with refiner, but only 7.91% wins with SD 1.5, so when I look at the size of the model, the SDXL base is 4.6x better in this popularity context while being ~3x larger, but with the refiner, the total is again 3x larger than the SDXL base, but the relative improvement in that popularity contest is just 30% more

so even when looking at their numbers, it's clear that if you search for tradeoff between number of parameters and quality, ignoring refiner is rational thing to do, because you just get too small quality improvement for large increase in compute complexity

[astral goblet]

theres a few of these human preference charts in sd3 paper too. i get it. it's really hard to evaluate models. It just seems to me that this is a really loose measure to base all evaluations on. It seems like fuzzy data to me

[vagrant fox]

those seems like cherrypicked numbers anyways

[tough lion]

I don't think it's cherrypicked, but I agree the evaluation is really fuzzy

[iron zenith]

so will the discord bot will turn on again with SD 3 ?

[tough lion]

so just now I am reading the SDXL turbo paper, and....they also compare it to the SDXL base, without refiner.

[lone hawk]

who knows)) all they do is giving us empty promises on reddit and showing the most boring portrait pictures on shitter, nothing really happening... no one got access to SD3.

[vagrant fox]

there is idea, rent out 1000s of gpus and make your own model

since you cant wait few weeks for free one

[iron zenith]

it's just for the easy access that discord bot gives us

[vagrant fox]

i was responding to meka not you

[iron zenith]

really good results without having to mess with the settings... will we ever get that back ?

[vagrant fox]

it's like saying i dont want to let machine know what i want it should read in my mind, settings are for fine tunning, whats the issue?

[iron zenith]

that for dumb people like me messing with the settings will results in artefacts and not good quality images

i'm not saying it isn't good

i'm saying that is isn't as magical as discord simplicity, you say something, and you have it

[rich kestrel]

use MJ for that

SD really isn't a one-click wonder

[iron zenith]

it was

[rich kestrel]

and thats a good thing

[iron zenith]

until they removed it

midjourney got only the newbie part

so we are limited with it

[lone hawk]

"few weeks" soon turning into a month 😄

[iron zenith]

before there wasn't any limitations

[lone hawk]

i don't want SD3 already... keep to it))

[iron zenith]

like the dumb people could create beautiful things on discord

[fervent thunder]

Is it worth it going for high steps?

[vagrant fox]

nothing like crying about free stuff needing more time to make it work like it should, you would also be one of those guys that rather have rushed game release and play with shit ton of bugs, right?

[iron zenith]

and the smart people could create even more beautiful things

[fervent thunder]

What is the difference really when you go to like 50 60

[iron zenith]

so why did they stop and told us nothin

[vagrant fox]

it dependso n complexity of art style you are trying to replicate i believe, anime and smooth stuff needs less steps than photorealism i think

[vagrant fox]

stop what? i always had to fine tune SD vs having to just spit out simple prompt in MJ

[fervent thunder]

So it does result in more detail?

[vagrant fox]

yes

[fervent thunder]

Is there such a thing as too many steps

That break the image

Like over 100?

200?

Nvm it goes up to 150

[vagrant fox]

i suppose both ends of spectrum could be damaging

[fervent thunder]

Im gonna play around with 50

[iron zenith]

the discord bot

[fervent thunder]

to 70

i always kept it on 20-25

[crude bay]

Some Loras seem worse with high steps

[iron zenith]

the discord bot permitted to do the same thing as MJ... ending in better results and for free

[vagrant fox]

recommend them option to donate to upkeep free hardware they offering for people that cant afford their hardware and then spam images there

[fervent thunder]

Which bot?

[vagrant fox]

if every person like u donated fiver a month maybe they keep it up 24/7

[fervent thunder]

Was there a bot equivalent to midjourney quality for free?

No way

[iron zenith]

yep

[fervent thunder]

Even in prompt simplicity?

[placid wasp]

my local install can go up to 999 I just foundf out by asking for 2000 steps

[fervent thunder]

Like writing 4 words and getting something really cool?

[keen rose]

I thought midjourney itself was a discord bot

[iron zenith]

yeah

[fervent thunder]

Then there must be a setting to increase the limit

[iron zenith]

better than midjourney

for free

[fervent thunder]

What was it called

[iron zenith]

DreamBot

[fervent thunder]

And now its disabled..?

[iron zenith]

yes

[fervent thunder]

Theyre making it paid or..?

[iron zenith]

and they didn't tell us about it

like

[vagrant fox]

didn;t told about what?

[iron zenith]

"it's in maintenance" "we close it forever"

[fervent thunder]

was it dreamboat by any chance?

not bot?

[placid wasp]

Mine is set in the Ksampler

[iron zenith]

the closing of Dreambot

[placid wasp]

im ComfyUI

[iron zenith]

it was a SD bot

[fervent thunder]

They probably saw how good it was

And decided to make it paid somewhere else

If it was made by a third party

[vagrant fox]

well whats the difference? don't expect free hardware to last forever, they will bring it back after sd 3 release most likely to gather more information about usage but what is the issue for u to run it locally or in google collab?

[iron zenith]

simplicity

[sinful sparrow]

SD bot only difference I'm guessing they mean it was all set up

[iron zenith]

the bot was finetuned

[sinful sparrow]

Vs doing it urself now to be local

[fervent thunder]

No

[iron zenith]

and just 4 words you could have something really good

[fervent thunder]

Difference is

[vagrant fox]

following this logic sd3 will be paid too, yikes, maybe they decided is not worth sharing resources with people spamming mostly dumb shit?

[fervent thunder]

In midjourney, if it was like this, you could type literally anything without much thought or finetuning

And get awesome looking images

[sinful sparrow]

Otherwise the image u saw in SD bot could be reproduce locally iirc, yea idk MJ

[fervent thunder]

Would take more time to replicate in SD

More knowledge

[iron zenith]

i mean it wasn't awesome but for simplicity there wasn't any major artefacts

my profile picture was made with this bot

in just one fricking line of prompt

[vagrant fox]

you can do the same with sd anywhere its not that bot was magical

[sinful sparrow]

All sd bot did was do the work for u

[iron zenith]

yes

[sinful sparrow]

U can indeed do it urself

[iron zenith]

I don't wanna pass 3 hours finetuning my bot just to remove an incorrect shadow on my drawing

like the point with this bot was simplicity

if they closed it because of the money, ok i totally understand that

but then tell us something else than "the bot is down until we are ready to return"

[keen rose]

so wait is Matendo having a problem because previous versions of SD are unavailable or something?

[iron zenith]

no

[vagrant fox]

because he cant use free sd bot on discord

[keen rose]

oh a discord bot

whatever then

[vagrant fox]

get SD on your pc and run it that way

[iron zenith]

i can't find this "magic tuning" anywhere

and i don't wanna pass 3 hours finetuning my models to just have a banal drawing

the staff won't say anything about the future of the bot

is it dead forever ?

we don't know

[vagrant fox]

i hope so

so people can learn how to use it locally 🙂

[iron zenith]

it takes so much TIME for BAD RESULTS

[vagrant fox]

its not like they put magical lora into bot

[iron zenith]

quite

i tested prompts that didn't make any sense

like "wearing glasses" and in the negative prompt "glasses"

[vagrant fox]

then you either badly constructed prompts or used wrong models/loras for what you wanted to achieve

[iron zenith]

the bot didn't gave me glitched glasses like the local version did

[true canopy]

im also having prompt issues

[iron zenith]

that's the point

[vagrant fox]

yikes, no shit it was wrong resoult if u ask it to make glasses and say not to draw glasses

what is your logic

[iron zenith]

you're there complaining "i have prompt issus"

[iron zenith]

it was a test

to establish the limits of the bot

Local version gave me glitched glasses, the bot gave me one image with glasses or no glasses at all

[iron zenith]

there was no "prompt issues" with the bot

[vagrant fox]

glitched glasses is more realistic resoult since you asked for two in same time in my opinion

[brittle maple]

i dont understand somebody can use for free sd3 now?

[vagrant fox]

bot just didnt do what u asked for

[iron zenith]

no

[vagrant fox]

yes

[true canopy]

im having trouble making anything manly haha

[iron zenith]

because the best version is the one that didn't have any artefacts

[vagrant fox]

you said in prompt wearing glasses and said in negative prompt glasses

what else it should produce?

xDDDD

[charred mesa]

https://x.com/Lykon4072/status/1766205651351478473

new SD3 image

[true canopy]

thats..almost scary

[brittle maple]

its work now for all?

[vagrant fox]

now ask him how many times he tried to produce this image before getting this quality

[vagrant fox]

no

[iron zenith]

it should ignore the mind of the creator and do one image or another

[iron zenith]

never giving me something glitchy

[brittle maple]

only for money?

[iron zenith]

never giving me artefacts

[vagrant fox]

or you shoulkd learn how to use prompts instead of confusing program intentionally

[lone hawk]

omg it can do single person portrait!!!

Lmao

[charred mesa]

yeah its magical

[lone hawk]

😄

[iron zenith]

they had a technology to make even the dumbest persons use this AI

[charred mesa]

I guess its just to show that its a good base model when it comes to aesthetics

[iron zenith]

they didn't sell it

[charred mesa]

we want more prompt adherence posts though :(

[iron zenith]

it's nowhere now even on clipdrop

what have they done with this magical technology

even with really bad and dumb prompts you would have something that's not glitchy

[wicked sonnet]

Are there some extensions you "must have"?

[vagrant fox]

dumbest people will keep generating waifus and memes only anyway, who cares about them, if you confusing prompts intentionally don\t be surprised sd don't know what you want

[iron zenith]

bot did

[vagrant fox]

then in logicall light bot didnt read your propmpts correctly and was confused and spit out 2 images instead of one

that reads prompt and negative prompt

[iron zenith]

yeah but the image was still coherent

[brittle maple]

rope,facefusion control net

[vagrant fox]

but was not what u asked for

[wicked sonnet]

what's rope?

[iron zenith]

like two twins one with glasses, the other not

[astral goblet]

for that, something like what foocus provides by default is what midjourney does. A gpt2 engine that extends prompts consistently so same seed makes the same results.

there's also "CADS" extensions for a1111 ui that shakes up the attention so that it doesn't fall into cliche patterns

[vagrant fox]

and both of them was not image you asked for still

[fervent thunder]

Well midjourney is still an option matendo

[brittle maple]

change face and adapt it for picture

[iron zenith]

i'm looking towards it yeah

[fervent thunder]

I should try that

[iron zenith]

i started on midjourney

compared it with discord SD bot

SD was better

[astral goblet]

on a1111 i use dynamic prompts for the same gpt2 prompt extending

[iron zenith]

i think i shall move back onto midjourney

[fervent thunder]

Is that an extension

[brittle maple]

mj bot vs sdxl?

[iron zenith]

SDXL bot

[astral goblet]

dynamic prompts is an extension. helps a lot. wildcards and procedural prompting

[wicked sonnet]

I suppose but I usually just generate anime which doesn't really need exact faces?

[fervent thunder]

I used to just manually ask the chat gpt bot lol

[wicked sonnet]

I'm searching for an extension that makes it easier viewing prompts

[astral goblet]

watch out on MJ. One user can crash the whole system

[brittle maple]

you can put some face in anime

[astral goblet]

very fragile

[wicked sonnet]

I suppose

I'll check it out later I think

[fervent thunder]

I dont get it

[wicked sonnet]

MJ?

[fervent thunder]

Wdym change face

Like fix it?

Edit it?

[wicked sonnet]

either that or deep fake?

lol

[astral goblet]

midjourny. you'll risk your whole organization getting bannt from the service forever if you use it to much and bring the whole server down

[wicked sonnet]

there's safety guards for that? and midjourney surely doesn't build a bridge with just one pillar lol

[astral goblet]

their bridge fell last week. and they blamed 2 stability accounts

"effective immeidately we're banning ALL stability related accounts"

[wicked sonnet]

maybe but unless their hardware exploded then no real damage in a sense?

[astral goblet]

you'd think someone like MJ would have safeguards from one or two accounts bringing down all their servers, but it turns out they don't. By their own admission

I'm just saying, you're paying for a service that is going to blame their users for it failing

[wicked sonnet]

Guess so but local is the way to go

[fervent thunder]

Yep

[keen rose]

I got a question in #🏞｜general-with-images, if anyone's willing to answer with some tips

[brittle maple]

what your queastion need extension for anime?

[wicked sonnet]

I want a way to visualize prompts similar to how tagcompletion does it

but live

[astral goblet]

#📣｜announcements dead channel 😦

[waxen vector]

when will SD3 be released?

[meager tinsel]

I'm not sure if it's allowed to talk about NovelAi, but I used Stable Diffusion for a year, and when it comes to inpainting in anime, it always takes many attempts and edits to remove or add something. I tried NAI, and in a few attempts, if not the first, it does what I want without going through a thousand configurations. And I'm someone looking for the result to be 100% anime, like a screenshot, not artistic. I noticed that inpainting is a completely different world between these two.

[past heart]

Is there a place for the script kiddies playing with the API & stiching videos together?

[astral goblet]

today according to emad's reddit comments

next week probably though

i don't think stability holds weekend office hours and they're closed in UK already

[spring karma]

hi guys i encountered problem how i fix it i press any key but its didnt happen anything, thank you for now
RuntimeError: Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check Press any key to continue . . .

[waxen vector]

thanks!

[astral goblet]

i woudln't thank me yet

[waxen vector]

but if he said it, he should do it. hehe

[warm imp]

Still looking for AI Horde related img2img help . Anyone familiar with it who might be able to help ?

[astral goblet]

It's emad. He says a lot

i'm feeling next wednesday. That'll probably be when bot access opens. just a strong feeling i got. probably wrong

could be longer

[sinful sparrow]

assuming ur talking about https://stablehorde.net/ this ai horde try asking in their community discord

[astral goblet]

thats an old domain. They renamed to https://aihorde.net/ because they do LLMs too

[sinful sparrow]

ohh ty yea they should have a discord there

[fervent thunder]

i can never get the images just right you know, always something wrong with it

[warm imp]

I left their discord bcus didn't feel comfortable staying

[sinful sparrow]

Unfortunately idk if anyone would know sometimes ppl from horde look here but no guarantee

[astral goblet]

well i mean.... that community are going to be the users of horde. they'll be most familiar with helping. Here you won't find horde users most likely.

When you're up shit creek, try not to throw your paddle to the shore.

[lone hawk]

they are scared to let people try it on discord, and you asking to release something!? Forget about it!! Lmao

[waxen vector]

never

[vagrant fox]

scared? i see you keep crying first about they shutting down bot, now that they need to take time to finetune and make sure new model is as good as it can be, get a grip

[lone hawk]

release it, i can finetune it myself

[vagrant fox]

do your own model then

[astral goblet]

^

" i want it alll. I want it ALL. I WANT IT ALLLLL. ANd I Want It NOW"

[vagrant fox]

YES I PAY NOTHING GIVE ME EVERYTHING NOW

yikes those people

[karmic cedar]

I think everyone seems to be overreacting to each other’s messages

[brittle maple]

The whole difference between them is described as follows: MJ is a street artist who draws for you for a little money. Fast, quite high quality, but it cannot be said that you can interfere with the process.

SD is not an artist, but a very smart set of tools. He won’t draw anything for you, but if he has the skill, effort and desire, he will draw exactly what you want.

[astral goblet]

an over reacting is bad mmkay

[karmic cedar]

The human is the input. The idea is the seed.

[astral goblet]

cause when people over react, reactions go overboard mmkay

[karmic cedar]

It’s true!

[astral goblet]

personally, i just think people need to stop blaming external factors for their own issues

[karmic cedar]

Absolutely they do

[vagrant fox]

and stop thinking that world is there to serve them

[karmic cedar]

Most folks take all of this stuff for granted as it is.

[astral goblet]

why can't the govenrment make laws to force stability to release it today?

fucking bs

[karmic cedar]

lol, they’re making plenty of laws. And most likely, none of them serve you all that well

[astral goblet]

i like the "don't steal" laws

they've done me pretty good

[karmic cedar]

It saddens me to think that they would enforce neutered models with handicapped creative flexibility on account of something like an aesthetic / style copyright. Because as more individuals weigh in on whether or not their work should be included, models will continue to transcend the individual such that the major powers will target style itself.

[buoyant moss]

Can't complain about free

I'm grateful they open source anything at all

[karmic cedar]

Ah, but that’s not the argument.

[vagrant fox]

and thats more attitude we need to see

[astral goblet]

imo, datasets should be culled.

they aren't taking the weights after they're trained and lobotomizing them. The models are fully developed and un neutared

the dataset is the genes, not the nutsac

[karmic cedar]

I understand.

My point remains ☺️

[buoyant moss]

My view is, if there wasn't Stability AI releasing models (neutered they may be, we'd likely have no open models at all, and just DALL-E/Gemini/MidJourney

So I'm just grateful they are willing to open source what they have

[astral goblet]

datasets that just leave it all in there, are tainted. the kind of taint after a long 5 day hike where you've had no showers.

you HAVE to cull datasets

[buoyant moss]

And if no NSFW by default is the price we have to pay for open models, then such is life

[charred mesa]

as long as anatomy isn't terrible or blood isn't culled as well I honestly don't care

[buoyant moss]

I am grateful for what is, instead of what may be

[astral goblet]

People still think SDXL is incapable of nsfw, because they can't use prompts from 1.5 porno models on sdxl base

Just train it in there. Don't be a newb

[buoyant moss]

ikr

[karmic cedar]

This is a matter of defining what style represents. Let’s say you have 51% artists of a certain range of style chime in and say that they’d rather their work not be included, while the other 49% does. Well, that’s enough data to invoke the style regardless.

[vagrant fox]

removing nfsw from image generator is still less harmful than biased and broken gpt's like gemini

[charred mesa]

if SD3 is only as censored as SDXL then we should be fine

[astral goblet]

The artists who were "removed" are all alive today. They're ALLLL derivative styles.

[karmic cedar]

As is your comment

[buoyant moss]

@alex just so you know, I did discuss this with Stability staff in person at NeurIPS in 2023 Dec

[astral goblet]

just prompt for an older artist. 51% of artists weren't removed

[charred mesa]

oh and what also makes me happier is that artists opted out and yet we still have good quality visuals

[buoyant moss]

They suggested that the CLIP model was just as important for style as the image dataset for training the diffusion model

[astral goblet]

you just called my comment derivative? What does that even mean? Do you have a point?

[vagrant fox]

most modern artist's are like SD themselves, getting inspired and create remix of whats been out there anyway

[buoyant moss]

Which is why they went with CLIP + Open CLIP in stable diffusion

[brittle maple]

I heard it is censored and some content is prohibited there

[karmic cedar]

So, with that in mind—does that resolve the inherent problem that these artists have? I fear it does not.

[astral goblet]

prohibited? you heard someone blowing smoke up your ass

[buoyant moss]

So going off what I heard in my discussion

They seem to suggest, a style can be not in the image training data, but if it was in the training data for CLIP, it can still help the full model reproduce a style

[brittle maple]

why agression

[slate dust]

I just wish fewer people would try to copy other artists and actually try to be unique artists themselves

[astral goblet]

it protects stability legally because they took reasonable measures to prevent their names being in the dataset.

[astral goblet]

its a figure of speech. They were putting you on. Trying to lightyou up

[buoyant moss]

My 2 cents--- better to have a neutered model than no model at all

[karmic cedar]

Stability could get knocked down much easier than OpenAI in a NYT-style lawsuit, so I’m just curious ☺️

[shell tendon]

Always going to be lazy copycats so matter what

[karmic cedar]

Sure.

[astral goblet]

neutared is the wrong word. They're not taking the finished weights and editing them. That would be quite a feat

[charred mesa]

yeah

[vagrant fox]

sadly u missing point that 99% of those artists are copacats and remixers themselves

[buoyant moss]

.... I do generative model research as part of my job

[karmic cedar]

No, but to suck the data from a training set is to make Swiss cheese.

[buoyant moss]

It is literally my full time job

[charred mesa]

I mean I guess some arstyles must have become oversaturated over a long time?

idk what to say

[karmic cedar]

In any case, I anticipate slightly less robust models going forward—but we’ll all be able to work around them with existing tools.

[astral goblet]

dunning kruger has demonstrated itself to me time and time again. many people who have a job, don't know what they're doing in that field still

calling models neutared tells me clearly that someone has peter principled their way into a position

[buoyant moss]

oof

[lone hawk]

on no... they showed a knight running from a dragon and it looks just like a Cascade... can't properly handle sword in hands, etc, we're doomed!!!

[karmic cedar]

You are very bite-y.

[buoyant moss]

ikr, he's a fiesty one

[brittle maple]

sdxl can generate faces at a distance without distortion?

[astral goblet]

gotta be against insidiousness

rawr

[slate dust]

It's a starting point, but artists that stand out have created their own styles

[buoyant moss]

But regardless

I'm kind of excited to see if v-prediction will improve the deep fried look stable diffusion images often have

[lone hawk]

look at the sword... https://twitter.com/Lykon4072/status/1766218735944118470/photo/2

[vagrant fox]

and what stops other people manually recreating them? it's not like someone has monopoly for art

[karmic cedar]

Me too!

[buoyant moss]

And if zero-SNR will improve the ability for SD3 to generate very bright & very dark images

[astral goblet]

2.1 has vprediction

[vagrant fox]

not an issue you can do inpaiting and revisions of this hand

[buoyant moss]

https://arxiv.org/abs/2305.08891

^ the zero-SNR paper

[karmic cedar]

I wonder if SD3 will have alpha support? Anyone know?

[brittle maple]

why 2 meter distance problem for ai)

[buoyant moss]

@alex I doubt it will out of the box

[karmic cedar]

Sure

Hopefully easy to make happen quickly ☺️ such a nice thing

[buoyant moss]

Reading the alpha diffusion paper, it suggests they need to re-train (a LoRA).

I agree

Would be nice to have

[karmic cedar]

Ahhh, that makes sense.

[astral goblet]

#🏞｜general-with-images message looking at sword

[true canopy]

after 3 days im starting to just barely understand what im trying to do

[lone hawk]

so bad... omg!

and i was thinking it's ready for us!

[buoyant moss]

But a problem is, a LoRA can't be easily re-used if the model is changed from epsilon prediction (SDXL, and SD1.x) to v-prediction

Even if the model architechture is the same

[astral goblet]

sd3 weights were probably being built before that research was integrated. we'll likely see a new set of weights sooner than later

[astral goblet]

i'm happy with it. i'll take those weights

[brittle maple]

I recently made a battle between sd and mj who can better draw a magical forest with a lake, no one won)

[lone hawk]

now i see why... it needs more training, Lol

[buoyant moss]

So in that sense, I think switching to v-prediction when they switch to diffusion transformer (from diffusion conv in SDXL/SD1.x) is the right decision

If they are going to break all loras, then now is the time

Might as well get better training stability while you're at it

[astral goblet]

prompt issue more than a model issue

[lone hawk]

no, don't release SD3 model in this sorry state!!!

[karmic cedar]

I’m really hoping we start seeing some other competitors in Stability’s league. Open source should not feel like a single destination point…

[brittle maple]

same prompt

[astral goblet]

prompt for old lady of the lake paintings. or "Legend" style magical forest scenes

[lone hawk]

i'm really against the release right now!

[buoyant moss]

@alex if you're willing to tolerate less realistic images, you can try Playground 2.5

[astral goblet]

that has no value

[vagrant fox]

seems like you can't make your mind up, first you rush to release now you are disgusted by current state, yikes

[buoyant moss]

https://huggingface.co/playgroundai/playground-v2.5-1024px-aesthetic

[astral goblet]

obvious troll is obvious

[buoyant moss]

@karmic cedar

[vagrant fox]

all people asking for quick release should be duped with access to 1.5 model

[astral goblet]

HAHAHAH

[buoyant moss]

The problem with playground v2.5, is the images are more obviously diffusion generated

[karmic cedar]

Thanks! I forgot about this one!

[astral goblet]

they'd fall for it 100%

[buoyant moss]

Super super high bokeh

[karmic cedar]

Ahhh yeah

That’ll come with time tho. Same as SD

[vagrant fox]

placebo would work on them for sure

[buoyant moss]

Like you know how sometimes you can just tell an image is diffusion generated? Playground cranks it up to an 11

[brittle maple]

sd like a genie if you say the right word it will make your wish come true

[karmic cedar]

Absolutely, yeah. It’s just where their code is at

[charred mesa]

This is what I worry about SD3. Will we be able to prompt it to not have bokeh?

[buoyant moss]

SDXL does have more boken than most images, Playground 2.5 is crazy at adding bokeh

[charred mesa]

I want narrow aperture images while it still looking good

[karmic cedar]

I wonder…if StarCoder 2 will come in handy for anyone building AI models rn 🤔

[lone hawk]

thank god for that dragon pic, it really opened my eyes! Lmao

[shell tendon]

Emad implied they know we hate bokeh

[buoyant moss]

Yeah, then the waxy images playground 2.5 generates is probably not for you

[astral goblet]

Easy to get no bokeh shots. People always obsess over blurred backgrounds. That's a prompting issue and they see what they want

[charred mesa]

good, thanks

[lone hawk]

SD3 isn't ready guys, not even close...

[astral goblet]

Bokeh looks good and people often promote those images

[astral goblet]

trollololololol

[charred mesa]

honestly the sword looks fine to me

it needs highresfix lol

[buoyant moss]

I think the bokeh in Playground v2.5 (and to a lesser extent SDXL) comes from the RLHF step

[charred mesa]

its trying to make that detail from a very small pixel density

its like a background face

or a hand

[shell tendon]

Bokeh is lazy and ruins most images iwo

[astral goblet]

doubt. SDXL had the bokeh artifact on the bots before RLHF happened.

Bokeh is just aesthetically popular

[buoyant moss]

People see images with bokeh, and they like it, and then the model generates pictures with more bokeh

[shell tendon]

Basically turns every image into a 1-3 subject image

[astral goblet]

selection bias is selection bias

[shell tendon]

With complementary lighting around them

[charred mesa]

the only upside is that it makes the background issues harder to notice, thanks to it all being blurred out lol

[astral goblet]

you believe something, then only see evidence towards that

[shell tendon]

Yup so it's a good way to cover up model problems

[charred mesa]

🤷‍♂️

[astral goblet]

we're talking about models with billions of parameters. It isn't stuck on bokeh.

[shell tendon]

Distant large objects look totally fucked cuz they're rendered small due to distance? Just blur it!

[charred mesa]

yeah

[buoyant moss]

whew flowwolf, you sure are a feisty one

[charred mesa]

Negative prompts are gonna be so powerful

[astral goblet]

eh

[charred mesa]

you can prompt whatever you want removed and it will be more precise about it

[buoyant moss]

I try to not engage with anti social behavior online

[brittle maple]

show andtell xl is like bot?

[astral goblet]

the fuck?

[buoyant moss]

^ like that

[charred mesa]

lol

[vagrant fox]

yikes

[trail lion]

So if we get more bokeh, and more delay on the release, that seems like the best compromise

[vagrant fox]

snowflake alert

[astral goblet]

"i got no good arguments so i'll call you antisocial"

smh

swear words hurted feels

[karmic cedar]

We are like wizards in training in a giant abstract wizardry school so everyone chill with their magic

[buoyant moss]

@trail lion I'm not against bokeh per-se, but if people keep voting on images with high bokeh we're going to end up with more bokeh

[karmic cedar]

(Also I talk like a goofball so feel free to direct your criticisms to that)

[astral goblet]

sdxl leaned to bokeh on the preview already, before RLHF had any sort of influence.

It's just popular in the data set and people tend to prompt towards photos more than other styles

your assumptions are not predicated by fact

[vagrant fox]

we all are goofballs trying to sound smart from time to time, feel like at home brother

[astral goblet]

I try my hardest not to comment about how people say things, and instead focus on what they're saying. While 'm not perfect here, i believe resorting to name claling and personal accusations is just... well its a huge cop out

[karmic cedar]

I just keep telling myself…”the text encoder provides the context”

[astral goblet]

"Prompting is King" is one of my mantras

[karmic cedar]

Good one

[astral goblet]

:king:

aww no worky i have no nitro no more

[karmic cedar]

I have yet to cross the nitro line myself lol

Maybe one day

[astral goblet]

wasn't worth it. higher upload sizes was convenient. about all really. i can live without mojers

[vagrant fox]

i got nitro because of free trial only, 10 usd a month for animated background and few emojis does not do it for me knowing people in different parts of world have nitro for 4 usd

[karmic cedar]

Hmmm

I think I need to use discord more.

lol

[vagrant fox]

but i bought avatar effect and miniprofile effect for permanent ownership just to support discord after years of usage

[karmic cedar]

Makes sense

[astral goblet]

My steam profile roars so i'm satiated

[karmic cedar]

Oh mine is ancient. Im happy

[astral goblet]

wanna flex steam profiles?!

[karmic cedar]

Haha “lidox”

[vagrant fox]

not much to flex at mine
https://steamcommunity.com/id/zieloneciastko/

[astral goblet]

AHAH

[karmic cedar]

You’ll see such treasures as…FTL, Half Life 2, Deus Ex

[astral goblet]

i win. 19 years service. 18.. pfffyftt

[karmic cedar]

Yep, I was a little late to the party

[astral goblet]

https://steamcommunity.com/id/scionoics/

ngl. Trusting steam wiht my CC back then was tough

i wanted physical discs for years

[vagrant fox]

same, i used to play at my friends pc since he had ethernet not me, i joined world wide web party super late

[karmic cedar]

Holy games Batman

[astral goblet]

i promised myself as a kid i would never want a game again. i'm about there i think

[karmic cedar]

Anyone here an FTL fan?

[vagrant fox]

thank steam for all shelf space u have now for useless stuff u collect instead of games lol

[charred mesa]

idk what Lykon means by this build doesn't enjoy this style of prompting

[untold herald]

Can I upload custom models in the Stable API?

[charred mesa]

I know that it's partially because of CogVLM, but did he also mean that it will improve over time?

[karmic cedar]

“That’s just like, your opinion, AI”

[buoyant moss]

@untold herald together.ai has dedicated GPU instance hosting

I'm not sure what their price is, I'd guess around $3.5 per hour for a 4090 class machine (or L40/A6000 Ada)

Maybe $4.5

[vagrant fox]

expensive

[buoyant moss]

Yeah...

[vagrant fox]

https://puzl.cloud/gpu-cloud A100 for 1.6 usd per hour

first google search

yikes 4090 for 4 usd per hour sounds like scam comparing to it

but i guess u pay for ready solution

[buoyant moss]

I'm not sure if the puzl instances are dedicated

But the prices you found on puzl are incredible

[vagrant fox]

well if you rent whole gpu then isn;t it fully dedicated to you?

[buoyant moss]

GPU maybe, CPU maybe not

[vagrant fox]

who would care about cpu while working on gpu anyway

[buoyant moss]

CPU -> GPU throughput can be important for some tasks

But even for server GPUs, power throttling is something that certain vendors do (although it is uncommon)

[vagrant fox]

i think a100 at 65% of its power would still eat 4090 all day long anyway

but im just guessing and know fuck all about dedicated servers

[buoyant moss]

depends on work load, if it is fp32/fp16 and if you use tensor cores or not

[vagrant fox]

you would use tensor cores for ai generation/ model creation i assume, why would you not?

[buoyant moss]

Not necessarily no

You need reduced precision to use tensor cores

*typically

At least recently Nvidia has been heavily marketing L40S (basically a 4090 with double the tensorcore speed for certain precisions) as a competitor to the A100, and they suggest in some work loads the L40S can beat the A100

[vagrant fox]

a100 has 6.9k cuda cores only lol so i suppose if you not use tensor cores and just vram and cuda it might underperform in certain scenarios comparing to 4090

[buoyant moss]

Although 4090 can beat H100 in fp32

CUDA fp32

Pretty nuts

https://github.com/openmm/openmm-org/blob/master/data/benchmarks.json

^ in pure compute task with no tensorcore usage, 4090 can exceed H100 in throughput

[charred mesa]

https://twitter.com/Lykon4072/status/1766239022735569070

[trim nymph]

going to sleep now, hopefully waking up to sd3 beta

[charred mesa]

same, fr

gn

[trim nymph]

sweet dreams!

[karmic cedar]

Night!

[charred mesa]

so it seems we are getting Comfyui support instantly

the devs are literally using it right now in comfyui

probably in a secret branch only they have access to

[astral goblet]

comfydev is an employee

[charred mesa]

exactly

[vagrant fox]

secret branch sounds like cool band name lol

[charred mesa]

Comfyui users: "Mark... This is good news"
Forge and Foocus users:

[astral goblet]

that'd be agood one to run during training. i been on ONI lately

[karmic cedar]

Yep yep, there’s also Stable Swarm which I haven’t tried yet

[charred mesa]

quality has massively improved, the eyes are no longer completely incorrect 💀

[astral goblet]

and then the 3rd group. People who aren't tied to one UI because misplaced brand loyalty

[charred mesa]

remember the older SD3 image of that woman

the eye was misplaced for example

[astral goblet]

the ginger?

[charred mesa]

yes

[shell tendon]

ugh gonna die waiting

[astral goblet]

might've just been a ginger thing

[charred mesa]

lmao

okay it wasn't that bad lol https://twitter.com/Lykon4072/status/1760688059744686104

[vagrant fox]

i bet prompt was "ginger teen on crack"

[charred mesa]

lmao

[astral goblet]

ayygh. no no. that eye is way low on one side. the forehead is abnormally huge. and also, red hair. that's just a personal problem though not an objective human deformity issue (or is it?)

point is, it's bad

[vagrant fox]

hahahaha

what ginger broke your heart

[astral goblet]

^ this guy ex gfs

[vagrant fox]

im not into gingers they crazy

[astral goblet]

yeah. yeah she was bro. way off the charts

[twilit cave]

the hype is chilling me

[vagrant fox]

you guys overhype it for yourselves don't expect massive revolution

models trained on base sd3 and loras for it will be worth more hype

[charred mesa]

the hype is killing me

[astral goblet]

it'll change everything. images will never be the same!

holgraphic generation WOOO WOOOO

[charred mesa]

true, but I'll still have fun prompting stupid stuff with the base model in the meantime

I prepared a big prompt list

[astral goblet]

we're going to have the FULL holodeck any day now

[charred mesa]

just for the testing phase

[astral goblet]

you want weird science? You got it

anything is possible with stablediffusionbo.com

[twilit cave]

lol I want to prompt it to debug python code but write the fixed code on a sign held by Squidward