🍥｜anime

[gloomy lodge]

DRAW YOURSELF NOW

[nova remnant]

Would eat

It's been a long time since I had a burgee

But I like pizza way more

idk that looks good now I'm hungry twice

[nova remnant]

Cute

Bun bun

[nova remnant]

https://tenor.com/view/small-small-center-zoolander-ben-stiller-ants-gif-5494197

[supple raptor]

Got tired of bing prompts getting rejected so I just prompted for the dog instead

[nova remnant]

I'm an AI generator expert, and I can confidently say that 0.9 is bad

👍

[wispy canopy]

10 imaginary points for the one who guesses which song inspired me to do this image

[nova remnant]

a wild guess.... spice girls - viva forever?

[wispy canopy]

not even close 😮

[nova remnant]

lol

it is close tho

just not that song 🙂

[wispy canopy]

haha

that's one wway to see it ^^

[nova remnant]

but speaking of Spice girls thats a good song unlike most others

[wispy canopy]

basically, ran the song title in llm prompt -> edited the whole resulting text only keeping the keywords, edited stuff to get the output coherent and more in flavour of what i wanted

[nova remnant]

ahh sounds interesting, i havent played with that

[wispy canopy]

llm prompting requires good coherent base models

as there's a lot of fluff words

[nova remnant]

i was thinking of running the new LLM from meta

[nova remnant]

but didnt get around to it

[wispy canopy]

llama2?

[nova remnant]

yeah think so

the one meta released publicly

[wispy canopy]

the finetunes are becoming pretty good

[nova remnant]

do you run it locally by any chance and is that able to access internet real time?

[wispy canopy]

the prompt was generated with OpenHermes-2-Mistral-7B-GGUF, for a 7b model, not too shabby

you run locally, and no, they cannot access internet

they're just "text prediction models"

[nova remnant]

oh crap, need the internet

thats a key aspect i want to have for a llm

[wispy canopy]

yeah, that's not available just yet

i'm pretty impressed that they've managed to make 7b models even capable enough to have conversations with

[nova remnant]

definitely and its been only 2 years about

[wispy canopy]

well, the OG llama was good at 33b, 13b was decent but sometimes just went haywire ... now the 7b models simply slap the 13b models of old around

which means conversational llm capability is certainly in reach of the commoner

[nova remnant]

didnt they also have mobile version to run with far less resources?

unless i read that wrong, but i think i heard something like that here in SD

[wispy canopy]

yeah, but realistically unusable

[nova remnant]

fruit posted something of that sort

oooh i see

[wispy canopy]

i'd have to read up to see if it changed

i think it was a 3b model (3 billion)

don't get me wrong, what it does is nothign short of amazing

[nova remnant]

i could be wrong but i think fruit mentioned about 13b model

[gusty palm]

oobabooga thing has a internet access extension thingy i think

[wispy canopy]

never checked it. could be fun

i also think in the early days, someone made a search-bot -> just get the headlines, and then summarize that

[nova remnant]

lol seriously thats a funny name, but yeah i want to try it out katie

[wispy canopy]

oobabooga = most used llmm text-generation frontend

[nova remnant]

ofc

found lots of guides on that ui

[gusty palm]

called superbooga or i think

[nova remnant]

lol ok

thanks for mentioning 🙂

[gusty palm]

i could be wrong having just reread the description
superbooga An extension that uses ChromaDB to create an arbitrarily large pseudocontext, taking as input text files, URLs, or pasted text. Based on https://github.com/kaiokendev/superbig.

[nova remnant]

the desc doesnt say accessing internet

Llama 2 says it cannot access the Internet, you can feed it links and ask it to summarize content or talk about it.v

LLaMA 2 is also not connected to the internet. That means it has a “knowledge cutoff” at December 2022.

[gusty palm]

theres one called autabooga but its no longer maintained

[nova remnant]

if bing chat and google bard didnt censor that would help

bing chat is using gpt4 but is heavier on censorship than bard

[gusty palm]

yeah best thing it can do is feed web contents as contexts due to above limitation

corporate ai censors ruin everything

[nova remnant]

yep but they have a reason to do that so that people dont use it with malicous intent, this problem runs deeper into the system that encourages malicious intnet to begin with

[wooden coral]

Still cursed but getting somewhere... maybe

[gusty palm]

woa

based on how easy it is to set up a no guidelines llm at home its just funny

i see why though

for the same reason they cant serve up nsfw content basically

[wispy canopy]

that's why finetunes exist

[gusty palm]

that reminds me i should go check for new 7b models

[wispy canopy]

TheBloke/OpenHermes-2-Mistral-7B-GGUF

seems pretty decent, actually

[gloomy scroll]

[nova remnant]

im assuming you are using restart sampler...

[slate narwhal]

no, I'm using dpmpp_3m_sde exponential, and trying a lot of things (unsuccesfully ) to change the model quality

[nova remnant]

whats that with such fine line detail ?

[wispy canopy]

`fairy logo, black and white vector art

Negative prompt: sad, sullen, swirl, blurry, error`

that's all

[nova remnant]

ahh ok, i think both of those sampler have that shiny look to them

[wispy canopy]

restart is pretty good, but there are some drawbacks at times

it simply sucks at upscale

and it's slow

GOD ITS SLOW

if you want it to do high quality

[nova remnant]

i played with it for a while but since i mostly use SD 1.5 models ive resorted to using DPM++ SDE Karras

[wispy canopy]

it's fast at 10 steps tho

and it doesn't get as clean lines as DPM SDE does

[wispy canopy]

used it as input for qr model

[nova remnant]

this is beautiful

like real photo of a flower

[wooden coral]

how to restore anime faces in a video, cause codeformer and all that is for realistic stuff and Face Detailer doesnt seem to work with multiple image inputs

[wispy canopy]

wouldn't adetailer work?

[wooden coral]

Im running comfy and the face detailer threw me an error

Codeformer works but not well enough for anime

Also DWPreprocessor is way better than openpose, but takes longer I feel

[nova remnant]

could be somewhat off, but this one talks about facial detail related stuff https://youtu.be/Da6J3wjBx3A

[wooden coral]

Ive heard of reactor will check it out thanks

[nova remnant]

np, btw roop is no longer maintained

[wooden coral]

Never used it

[wispy canopy]

why my fascination with fairies today? because of this absolute banger https://www.youtube.com/watch?v=nlGMXDQ5W20 (went to their live, they just as good as on album)

[nova remnant]

that music... love it

[wispy canopy]

they used that song as closing song

such a fckn banger to end the concert with

[nova remnant]

can imagine the climax there listening to it in a huge crowd

since you dig techno music, i'd recommend listening to hana

she has youtube channel 'hana'.. her music rocks

and she is hot

combination of sugar and spice

[wispy canopy]

actually, i'm not that into techno 🙂

mostly into melodic/high energy music

[nova remnant]

yeah thats it

give her music a try

melodic, high energy

[wispy canopy]

mostly got some techno beats, those won't work for me 😦

and don't get me started on my pachelbell rant

[nova remnant]

listen to her already

🙂

if her music doesnt get you moving youre missing some pulse lol

[wooden coral]

I dont even know what went wrong

Or where all the red stuff comes from

[nova remnant]

i like it, fun animation

[wispy canopy]

honestly -> leaves me ice cold XD

[nova remnant]

and that animation kinda looked good with the music im listening to by hana

[nova remnant]

you mean her music dont match your taste?

[wooden coral]

Trying to get a clean look

[wispy canopy]

exactly. it's not bad in any stretch

just not my cup of tea

[nova remnant]

oh wow.. but i dont understand

[wispy canopy]

https://www.youtube.com/watch?v=19zPEx6ZGvs

[nova remnant]

i mean you liked that techno music that you shared

[wispy canopy]

this is what i meant with high energy

it has techno influence -> but at it's heart it's an orchestral piece with not often used chord progressions

[nova remnant]

i like this one too but i have issues with too much beats

[wispy canopy]

my music taste is weird

oh. you woulnd't like the stuff that i can listen too then 😛

i mean... miracle sympony x by usao is... a whole different beast even

i'm not even gonna bother linking it... only seek that one out if you're really desperate for ... beats

[nova remnant]

i can listen to metal and hard rock too

but i prefer melodic progressive music more

[wispy canopy]

if you're into melodic instrumental, give arforest a try

that dude has some incredible melodic lines

[nova remnant]

ok

im still listening to the one you just shared

this has nice melody as well

kinda fusion

[wispy canopy]

not often someone likes fear and loathing in las vegas

[nova remnant]

is it?

[wispy canopy]

they themselves call it electronicore

[nova remnant]

its got thrash metal, melody all mixed together

[wispy canopy]

combition of hard rock/metal fused with synths

[nova remnant]

yeah i can dig it, i love that kinda mix

[wispy canopy]

it's a pretty bizarre combo, but i really like it actually

i got to know them because of the anime parasyte

the opening stunned me, looked up the band, incredible stuff

and got stuck on 焔 - sennzai again

[wooden coral]

getting the mentioned import error here but the fix doesnt work

[wispy canopy]

https://tenor.com/view/thats-the-neat-part-you-dont-invincible-gif-27194608

[nova remnant]

how do you like psychedelics?

[wispy canopy]

psychedelic songs?

[nova remnant]

yeahg

[wispy canopy]

not really into those

[nova remnant]

instrumental tho

what do you think of this https://youtu.be/pCflIwFjiF4

@wispy canopy you're back from the dead

[wispy canopy]

was grinding out some achievements on poe

[nova remnant]

[wispy canopy]

i think i'm on 30 now

[lyric zealot]

for some reason, middle is the best

[nova remnant]

2m Karras is very nice

[nova remnant]

Wait so I'm not just crazy and restart really had the more bright look on things

used to use 2m karras mostly then i switched to sde karras, to me that made a good difference with image quality

[nova remnant]

Could depend on model

[lyric zealot]

sde karras only? not the ++2m sde karras?

[lyric zealot]

[nova remnant]

yes just the dpm++ sde karras

[lyric zealot]

might be how i implemented the workflow

[nova remnant]

but i have to look close to the difference with 2m models

but im certain sde karras made a good difference from 2m karras

and you know what that could be related to model too, so far i have good results with using it

that music i shared is boring me to death after listening to high energy ones

@nova remnant how many steps do you use with SDE karras

30

plus 10 for highres

Hmm

[nova remnant]

i read on huggignface you can only notice visible differences with steps going as high as 50 max

[wispy canopy]

SDE are ancestrals if i'm not mistaken i stand corrected

[nova remnant]

no

[lyric zealot]

restart really brings in the refiner model effect

[wispy canopy]

they aren't?

[nova remnant]

dpm++ sde karras isnt

[wispy canopy]

then i'm mistaken ^^

[nova remnant]

usually all ancestral would have _a in it

you might not be mistaken but let me check to be sure

what do you know

[nova remnant]

Does the hires restart

thats ancestral and you are correct

Oh wait

It's the refiner

That explains it

DPM++ SDE Karras is an ancestral sampler. Ancestral samplers are a type of sampler that uses the noise from previous steps to generate the current step. This can lead to smoother images and better convergence, but it also means that the images are not as stable with respect to changes in the number of steps.

[lyric zealot]

i believe it's the refiner

since im using expmixlinev3 as refiner

[nova remnant]

That explains many things

[nova remnant]

Can you do refiner at 0.6

im glad you brought this up @nova remnant

With 2m Karras and restart plot

im lookin at 2m karras in new lights

Lmao

It's very nice

[lyric zealot]

it's cooking

[nova remnant]

Very stable with quality

[lyric zealot]

well

it's really fast

and good quality

so

[nova remnant]

@nova remnant ever tried ddim ?

[nova remnant]

yes

ddim was pretty good at low setings

[wooden coral]

Found the problem, running with reactor now, lets see

[nova remnant]

perfect for someone

i used ddim a lot on my 4gb card few months ago

[wooden coral]

[nova remnant]

and with steps as low as 20

[nova remnant]

Give

[wooden coral]

I actually got 2

[nova remnant]

But why

[wooden coral]

One for work and one for gaming PC

[wispy canopy]

ah, so i was right 😮

[wooden coral]

Got the gigabyte OC for work and the MSI Liquid Suprime X for games

[nova remnant]

Hmm yes

The big game changer, names on cards

[lyric zealot]

0.6 refiner

[nova remnant]

[wooden coral]

Home PC

[nova remnant]

Aquarium PC

[wooden coral]

Hyte Y60 best case

But now we get a Y70 with touch screen

[wispy canopy]

nevermind me shamelessly lyricsdumping 😄

[nova remnant]

Why does restart have the refiner model look so much

[wispy canopy]

bonus image ❤️

[nova remnant]

Does it go back from 100% after refiner kicked in

[wispy canopy]

restart and refiner happen quite close to each other

[wooden coral]

Whats the restart thing btw

[lyric zealot]

probably because within the 10 hires step, restart sampler does it's thing and restart some of the progress?

[wooden coral]

Looks neat

[wispy canopy]

restart happens around 70-85% (once it gets close to converging), and refiner usually at 80%

[lyric zealot]

[nova remnant]

But you're not using the refiner model on hires

[wispy canopy]

restart sampler -> specific to a1111

[lyric zealot]

oh true

[nova remnant]

Hmm

[wispy canopy]

you can get custom node for comfy

[nova remnant]

Eface knows the stuff

[wooden coral]

Ah well a1111 exclusive

[lyric zealot]

maybe cos im doing 30 steps

[wispy canopy]

but it's a lot more config work than the out of the box restart sampler in a1111

[wispy canopy]

https://github.com/ssitu/ComfyUI_restart_sampling

[wooden coral]

Oh is there

[lyric zealot]

and 0.6 refiner means, 18 steps of refiner model?

[wooden coral]

[nova remnant]

Does restart, restart on hires too?

[wispy canopy]

advanced restart, ddim (ddim_simple), restart(exponential) gives me best quality by default

[wispy canopy]

yes, and it prevents clean lines usually

[nova remnant]

Interesting

[wooden coral]

Gotta read into the segment stuff

[wispy canopy]

@nova remnant i usually use dpm++ SDE for the highres pass

[nova remnant]

Is there a sampler that's good for those

[wispy canopy]

haha, answered the question before it was asked!

[nova remnant]

Hmm

But overall

Not just on hires

[wispy canopy]

i went from dpm++2m to uni_pc to restart, just love sticking to deterministic

[wooden coral]

Oof reactor gave me a corrupted video output, gonna retry with a different setup

[nova remnant]

Something is going to explode

[wispy canopy]

@wooden coral you got a mini itx case?

[wooden coral]

Nope thats a full sized case, actually really huge

[wispy canopy]

ah, midtower

it looks small 😮

[nova remnant]

Mini or 4090

Choose one

[wooden coral]

https://hyte.com/y70touch?utm_campaigns=y70touch
Eyeballing this

[wispy canopy]

damn you're not wrong, that's actually a sizable unit

[wooden coral]

You can fit like a 25 cm thick rad in the side

[wispy canopy]

i'm into the sleeper camp -> antec p110 silent

[wooden coral]

I like my shiny tech

[wispy canopy]

i like it silent and cool

uhh... yeah, when SD or games are active, it's still pretty decently loud with all fans going all-out

[wooden coral]

I mean my system uses 2 AIOs so its silent and cool too

[wispy canopy]

got your gpu on aio?

[wooden coral]

Yeah MSI Liquid 4090

thing rarely goes above 45 degrees while gaming

[wispy canopy]

i was looking into a custom waterloop for both gpu and cpu

but that costs €€€

mine is around 73 max, courtesy of fancontrol

when the gpu gets active, cpu fan and case exhausts get active too, making sure enough cool air gets in

[nova remnant]

@wispy canopy I'm at 34, ritual is horrible I need help

[wispy canopy]

the reroll challenge?

[nova remnant]

Yes

[wispy canopy]

i had 50 done... and i think 45 were done on my fireball miner -> that build instaclears rituals

[nova remnant]

I sleep all day

[wispy canopy]

i should've bought enough comprehensive reports earlier now they're crazy fckn expensive -> still need to do doryani's

[nova remnant]

Ain't got time to map

[wooden coral]

Custom Waterloop is about 1000 bucks extra

AIOs are cheaper

[wispy canopy]

yeah, that's the price i also got

problem being, i need about 5x120mm to cool whatever i have now

[wooden coral]

Thats what I got too as AIOs

a 360 and a 240

[wispy canopy]

360 for gpu and 240 for cpu?

[wooden coral]

the other way around since I got a 13700k and shit gets hot af

[wispy canopy]

songmaster lyricsprompting + region control => fun unpredicable results

[wooden coral]

also the MSI card only comes with 240 Rad

[wispy canopy]

don't get me started on the mech with .. uhh... strange beams

[wooden coral]

Which is enough tho

[wispy canopy]

looked for aio's for 3090 fe

not a whole lot...

[wooden coral]

Laserdick

Yeah adding the AIO later sucks

Cause all cards are slightly different

[wispy canopy]

msi aio €2k? 😮

damn

[wooden coral]

I paid 1700 iirc, got a really cheap deal

But any other 4090 in the Y60 needs a vented sidepanel, which I did not like

Y70 will be thicker and doesnt need that it seems

[wispy canopy]

double 4090?!

[wooden coral]

No lmao

[wispy canopy]

oh, any other

yeah, because of the size

[wooden coral]

Yeah the Y60 wasnt made with 4 slot cards in mind

[wispy canopy]

i think not many cases are XD

4090 is a behemoth of a card

well, the cooler is

[wooden coral]

Thats why I got the liquid one

The card itself is really small

[wispy canopy]

yeah, it is

the bulk is all cooling solution

[wooden coral]

Its a hybrid card too, the other liquid 4090 is fully cooled by the aio and has a 360 rad

[wispy canopy]

@wooden coral may our fps be high, and temps be low!

[wooden coral]

I really shouldnt do a 120 frames animation for testing purposes

[wispy canopy]

... 'testing'

[wooden coral]

doesnt seem to work

maybe cause anime

[wispy canopy]

always anime XD

[nova remnant]

differences are so subtle

[lyric zealot]

yup

[nova remnant]

prompting dyanmic pose can make character poses come out strangely interesting

[wooden coral]

Dunno where the red comes from, I even put it into the neg

[nova remnant]

there is a lot of red for not expecting red to come out lol

i mean if you didnt have single mention of red then that's crazy

[nova remnant]

Dynamic pose:1.5

[nova remnant]

4xUltra sharp

[nova remnant]

i havent used weight on them ... just putting dynamic prompt made my characters take weird forms lol

[wooden coral]

Yeah trying to figure out where it's coming from

Cause prompt doesn't include red and I don't feed it an image besides the controlnet

[nova remnant]

It's funny

If you go too high like :2 they just want to come out of the screen lmao

I'm being serious

Lmao

not dynamic but they look cute

[lyric zealot]

is that good?

[nova remnant]

Try it

I think it's great

For the poor amount I can use hires

Actually I use it to tile

Same thing

[nova remnant]

I'm drawn to brighter things

Ultra sharp is brighter lmao

Or less contrasty

[wheat falcon]

Why so angry 😮

[lyric zealot]

true, the 4x-fatal is the darkest

[lyric zealot]

i'm always angry

[wheat falcon]

Remacri is darker I think

[lyric zealot]

https://tenor.com/view/shinobu-kocho-gif-24612346

i just restarted my webui

[wheat falcon]

Are the faces inpainted ?

[nova remnant]

used adetailer i always use adetailer by default

[wheat falcon]

Ah okay looks good

Which face yolo model did you used ? Is there a big difference ?

[nova remnant]

adetailer made things lot more fun and nicer for image gen 🙂

with all the default settings

[wheat falcon]

Ah okay

[nova remnant]

i used to tweak a few settings but with newer version no tweaking is necessary for regular images

[wheat falcon]

Ah need to try it again then

The faces looking really good

[nova remnant]

yeah, i get 0 defect on faces with adetailer

only issue i still have is with fingers

[lyric zealot]

fatal is darker

[wheat falcon]

Okay fatal wins

[lyric zealot]

or is this not what you meant?

[wheat falcon]

Add lolypop xD

[lyric zealot]

and she's smilinng for you now

[nova remnant]

how are you guys trying to assess the difference just from sight?

[lyric zealot]

is that a upscaler?

[wheat falcon]

YandereNeo XL is what I often used

[nova remnant]

i have no idea how you are spotting any visible differences from that set of images

[lyric zealot]

in my case, i just open up the image, and press left and right arrow key

[wheat falcon]

NKMD Siax 200k is good too

[lyric zealot]

like this ^

[nova remnant]

ahhh i see what you mean

[wheat falcon]

Check the details

[lyric zealot]

is that for XL only? Since there's the xl name

[wheat falcon]

Nope works on 1.5 too

[lyric zealot]

and i can't find that upscaler

[nova remnant]

now that you just explained the method i see some big difference

[wheat falcon]

https://openmodeldb.info/models/4x-Lollypop

With 2 l my bad ^^

[nova remnant]

curios to know .... any reason why you might be needing anything besides these?

[lyric zealot]

I mean, different quality of upscaling?

and also, like you saw there, some are darker and stuff

unless that's not what you meant

[nova remnant]

well 4x lollypop is ESRGAN 4x+ that comes with A1111

[wheat falcon]

For realistic images for example I prefer 4x ultrasharp if I don't use latent ones

[wheat falcon]

Its based on esrgan, its not the same

[nova remnant]

with added changes to tone, contrast besides sizes?

[wheat falcon]

Creators made them for different purposes

Some can do better pixelart for example or manga or realism

[nova remnant]

ahh i get it now that makes sense

[wheat falcon]

The esrgan ones in auto1111 are base "universal" ones

Anime6b for anime of course

[nova remnant]

i usually use either of those two, the images above i posted had highres fix with R-ESRGAN 4x+

[wheat falcon]

Which sampler did you used ?

[nova remnant]

dpm++ 2m karras with 30 steps and highres 10 steps

[wheat falcon]

Okay, can't decide between 2m and sde karras

[nova remnant]

i looked up the detail, 2m karras is better with quality and performance

[wheat falcon]

Okay, I need to make some XYZ tests then again xD

[nova remnant]

you can do that too to be certain

i used google searches and read up on the technical applications

[lyric zealot]

i did some previously

[nova remnant]

not really anime... but this is an example of how good Adetailer is even with realism, and this image also have a very small facial dimension but came out looking very aesthetic

[wheat falcon]

And what's your thoughts about them ?

[lyric zealot]

Im liking lollypop

[wheat falcon]

I mean between the samplers ^^

I tried 2M SDE Karras a lot, liked it

[lyric zealot]

ahhh, im really liking DPM++ 2M Karras, but if you are using refiner model, and want the refiner model to have more effect, use Restart

[wheat falcon]

Yea its a big difference there it seems

Have to play around with that when I'm at home

[lyric zealot]

yup, that's really the effect of my refiner model so

[nova remnant]

do i download the lollypop into StableDiffusion\stable-diffusion-webui\models ?

[lyric zealot]

Im still really liking how fatal is darker

[wheat falcon]

models/esrgan

[lyric zealot]

but lollypop is great for brighter upscaling

https://www.reddit.com/r/StableDiffusion/comments/179uuz3/introducing_stable_fast_an_ultra_lightweight/

i love how the first comment is always - "When auto1111?"

[wheat falcon]

Haha xD

True

[lyric zealot]

[wheat falcon]

Yep but the other two have cleaner details

[lyric zealot]

yup

so i guess it really depends on what you want

[wheat falcon]

True, "when Upscaler Merge?!"

[nova remnant]

ESRGAN 4x+ is generally better at preserving fine details and textures. It is also faster than R-ESRGAN 4x+. However, ESRGAN 4x+ can sometimes introduce artifacts into images, especially when upscaling by a large factor.

R-ESRGAN 4x+ is better at reducing noise and grain from images. It is also better at restoring images that have been damaged or compressed. However, R-ESRGAN 4x+ can sometimes blur fine details and textures. It is also slower than ESRGAN 4x+.

In general, R-ESRGAN 4x+ is the better choice for upscaling realistic photo images. ESRGAN 4x+ is the better choice for upscaling images that contain fine details and textures, such as anime and cartoons.

so lollypop is built to handle both real and anime images

[wheat falcon]

Yes, there are a lot of multi purpose upscalers. Remacri is also one

NMKD made a few for different stuff and an 8x Upscaler only for faces

[nova remnant]

using lollypop

[wheat falcon]

Looks good

[gritty spruce]

coming late to this discussion but I see no major differences between 2M and 2M sde with karras scheduler

they look almost exactly the same, don't know about efficiency or memory etc

[nova remnant]

yeah but we were talking about 2m karras and sde karras

[gritty spruce]

a lot of people who do realistic like best the 2M sampler with DDIM:uniform scheduler

I personally like karras better

[nova remnant]

sde karras is ancestral which 2m karass isn't and they have other technical diferences in features

i started out with 2m karras then swtiched to sde karras, now i just switched back to 2m karras

[gritty spruce]

yeah that's basically my preferred setup 2M karras

3M sde also works fine with karras, not with ddim

[nova remnant]

ddim was pretty cool to use with low setting when i had 4gb card

ever rendered decent images with 15 steps only using ddim

[nova remnant]

DDIM 15 steps no highres fix

[lyric zealot]

how long does it take to run 150 steps for 4090 users

[nova remnant]

ddim 150 in 57sec

using rtx 3060

[lyric zealot]

is 4090 just like a blink of an eye

[nova remnant]

maybe not a blink of an eye but definitely a beast if you get the 4090

[wooden coral]

Not really, just a lot faster

If you need a 4090 time for any task let me know and i can run it

[nova remnant]

i used sd on google collab that runs on tesla monsters cards but wasnt describable as blink of an eye

smol sus i think im achieving some of your cuteness signature with these images

[lyric zealot]

even i have no idea what my cuteness signature is

[nova remnant]

its you

your mental power

[lyric zealot]

https://tenor.com/view/blushing-anime-girl-hides-hide-face-hands-over-face-sneak-peek-gif-24165828

[nova remnant]

lol

[lyric zealot]

?

[nova remnant]

cute

but cuter w/o sunglasses

[lyric zealot]

true, but it's bright out there

[nova remnant]

the sunlight has no option but to be gentle on her cute eyes

[wooden coral]

IPA still works without issues
I wonder whats going on

[lyric zealot]

https://www.nvidia.com/en-us/geforce/news/game-ready-driver-dlss-3-naraka-vermintide-rtx-vsr/

it's not just about game, they talked about 2x faster speed for SD

[wooden coral]

Ill download and check it out

[lyric zealot]

please do

im too scared that my stuff will break

[wooden coral]

https://tenor.com/view/homie-shower-gotchu-sneaky-gif-13813267

[lyric zealot]

it's using tensorRT

[wooden coral]

I dont

[nova remnant]

is that applicable for all rtx cards?

[wooden coral]

Probably? Maybe its only for 4000 series idk

[nova remnant]

oh shit, i have 3060 not 4k sereies

[lyric zealot]

https://github.com/NVIDIA/Stable-Diffusion-WebUI-TensorRT

they said to download this extension

[nova remnant]

holly crap

this is some leap

[bitter latch]

hmm interesting

[lyric zealot]

sorry for linking so many stuff, but here https://nvidia.custhelp.com/app/answers/detail/a_id/5487

[nova remnant]

no dont be sorry at all

im checking all those links

[lyric zealot]

[nova remnant]

Minimum Requirements
GPU NVIDIA RTX GPUs with 8GB of VRAM
RAM 16GB RAM
Connection Internet connectivity during installation
Driver NVIDIA Studio Driver 537.58, Game Ready Driver 537.58, NVIDIA RTX Enterprise Driver 537.58, and above

[lyric zealot]

oaky god bless

im updating my driver

[nova remnant]

me too

[lyric zealot]

[wooden coral]

Thats only for A1111?

[lyric zealot]

there's only guide for A1111

[wooden coral]

Comfy times seem to be the exact same with the update

[lyric zealot]

[wooden coral]

Just a matter of time till comfy supports it id guess

[lyric zealot]

definitely

comfyui usually get things updated pretty quickly

[wooden coral]

Looking forward to people testing this and getting actual numbers, I take all Nvidia claims with a grain of salt

[lyric zealot]

already have problem downloading the extension, but that's because im dumb

[wooden coral]

But its crazy to me that Nvidia seems to have choosen to officially support A1111 kinda

[lyric zealot]

true actually

[wooden coral]

Or does that only look like big news to me

[lyric zealot]

it would make more sense for them to build their own app or something

[nova remnant]

crazy as in exciting or bad?

[wooden coral]

I dont think its bad per se, but Id wish for them to support a wider variety of UIs

Its a good step id say

[nova remnant]

yeah this will get more widely applied

[wooden coral]

https://nvidia.custhelp.com/app/answers/detail/a_id/5487
They even made a guide

[nova remnant]

but im kinda glad a1111 is supported

to begin with

[wooden coral]

Which was already posted mb

[bitter latch]

ok lets see if this tensor thing works

[lyric zealot]

seems like a full restart is needed after installing extension?

[wooden coral]

whats wrong with my video input setup

[wooden coral]

usually yes

[lyric zealot]

looks good to me, what you mean?

[nova remnant]

Users can specify quality levels between 1 (lowest) to 4 (highest). Higher quality levels will result in sharper images and improved artifact reduction. Higher quality levels require additional GPU resources, so if users need more horsepower to run creative apps or games while playing video, the quality level can be adjusted. All GeForce RTX GPUs are able to comfortably upscale using quality level 1, and xx70 class or higher are able to play most content at quality level 4.

any suggestion on that value between 1-4?

by default its set to 1

oh https://nvidia.custhelp.com/app/answers/detail/a_id/5448

[wooden coral]

Thats the video upscaling

Has nothing to do with SD afaik

[nova remnant]

right its for super resolution

just thought i'd ask it

[lyric zealot]

so from what i can tell, it's basically to prepare your GPU for a certain type of generation? So if you generate within these parameters, you will get boosted speed?

that's kinda like my layman explaination, if that's even correct

okay my cmd is definitely doing some stuff, not sure what tho

[wooden coral]

[lyric zealot]

so it's per model?

[wooden coral]

no clue

[lyric zealot]

okay it's done?

[gritty spruce]

if that is an error it looks really interesting, i'd keep experimenting with that bro!

[bitter latch]

yeah getting nothin but errors

[nova remnant]

??

[lyric zealot]

you can't use this with hires fix?

[nova remnant]

can you post the errors?

[lyric zealot]

i got error with hires

but it worked without hires

and it's fast

[nova remnant]

hmm .. im nearly done downloading

[bitter latch]

there is no ---> Click on the “Generate Default Engines” button.

[lyric zealot]

this on a 3070

[bitter latch]

wtf

[lyric zealot]

export

[bitter latch]

so got a feeling thats where it's messing up

what? explain

[lyric zealot]

export instead of generate

[wooden coral]

#🍥｜anime message
Thats how the IPA output looks, and the other one is with a video for openpose as input... Idk whats wrong

[bitter latch]

ok i'm not even getting that far

so meh

[gritty spruce]

are you joking? what size? I dont get that on a 4090

[lyric zealot]

[nova remnant]

eeesh i have to cut all the lan connection, getting slow d/load

[bitter latch]

oh well. screw it

[lyric zealot]

512x512

okay, the size that you generated the profile

it takes hires fix as well

so example

if you gen 512x512, hires fix at 1.5 times, ending res will be 768x768

then that works

[nova remnant]

a relevant question to this smol sus, what's more ideal to gen at 512 then high rex fix it by 2x or gen at 768x and highres fix it by 1.5x?

[lyric zealot]

you can't have higher res than 768x768 with this method

or at least i don't know how to yet

[nova remnant]

thats odd

[lyric zealot]

you can't hires fix and go over 768

okay, now i can do batch size 4

wait you can

there's the default settings place

i mean advanced settings

[nova remnant]

you can go over 768 then?

[lyric zealot]

yup

[nova remnant]

ahh nice

[lyric zealot]

you need to build a profile for that

i guess?

I'm gonna be honest, im just speaking without knowledge tho

[nova remnant]

im still stuck at downloading the extension

[lyric zealot]

good luck

[nova remnant]

not sure if things broke

does it supposed to show progress bar in console when download that extension?

[lyric zealot]

nope

[nova remnant]

cuase i got nothing on console for it

[lyric zealot]

that's why i also thought it bugged

but just wait for webui to be done

[nova remnant]

ok fingers crossed i will wait

[lyric zealot]

or you can go to your extensions folder and see if a new folder is created

[nova remnant]

oh thats clever, yes i see it working

[lyric zealot]

this is some interesting stuff

okay

for each lora, you also need to build a profile?

[dawn jasper]

That looks like Rin Tohsaka.

[wooden coral]

maybe because it is, made that a while ago and used it as base

[dawn jasper]

Her hair is Beautiful because it's Magical and Magical because it is Beautiful.

[nova remnant]

ok im done with intalling ext

now to get into the detail

[gritty spruce]

Toyxyz is inpainting blinks with animateDiff https://x.com/toyxyz3/status/1714290419570442496?s=46&t=0WU4EHlrcScluYAy-xmVqQ

[nova remnant]

i got some nasty python error with popup window

[lyric zealot]

[nova remnant]

cant be related to ext installation

i think i will relaunch webui

[lyric zealot]

[lyric zealot]

[nova remnant]

GS is not installed! Installing...
Installing protobuf
Installing onnx-graphsurgeon
UI Config not initialized

[lyric zealot]

i don't know if it's because it's in my ddrive

the loading of profile outweighs the speed increase in generation

[wooden coral]

Still tripping

[nova remnant]

something got messed up in my venv folder

but even with that errror prompt sd launches in browser with no other errors showing up in the page

im guessing the dynamic link for tensort rt isn't working when running the bat

cs1 and aryetis will have busy hours the next few days

[lyric zealot]

[nova remnant]

lol i deleted the venv folder and relaunched the bat to download the necessary few gigs of file now

[bitter latch]

yep thtas what I got so I just deleted it

[nova remnant]

deleted what?

[bitter latch]

mh would rather explore space in starfield anyhow

the tensor thingy

[nova remnant]

lol why.. ther has to be a work around

[bitter latch]

I only got 4gig vram so why bother. works fine the way it is. most of the time

[nova remnant]

the old settings didnt really connect with dynamic link settings

4gig on rtx card?

this update is for rtx card only

rtx gpus...

[lyric zealot]

the audacity to put 3070 as high-end

[nova remnant]

though mine is rtx 3060 with 12gb vram

3070 is good 🙂

[lyric zealot]

3070 is good for nothing

[nova remnant]

3070 is probably the right card for cute images

im guessing that's Ai sweet spot

[lyric zealot]

lmao

[nova remnant]

wait what is this? driver's update?

yeah and sd update

[lyric zealot]

this in 14.2

but no hires fix

I bet im getting worst speed with hires fix and adetailer is because of my sd living in ddrive

[nova remnant]

oooooh... is that regular drive or ssd?

[lyric zealot]

hdd

that's why the loading of profile takes time??

imma try something

[nova remnant]

you'll have to help me later, am out rn

oh, that's gonna slow down a bit

[lyric zealot]

but i have no idea how to delete profile

[nova remnant]

cs1o knows i bet

prolly to do with config.json and ui-config.json but better to check with him first

[lyric zealot]

it's definitely this

each profile is about 1.7gb

[nova remnant]

oh ouch

[lyric zealot]

[stable crane]

😄

[stable crane]

cute, 14.2 what?

[lyric zealot]

14.2 seconds

[dusk seal]

pov: the last thing rescue team finds

[bitter latch]

ummm

[lyric zealot]

discord

[gusty palm]

is there a textgen-ui-oobagooba discord?

im having like a unique problem that i can describe but

imna goto offtopic to post it so i dont spam up here

[lyric zealot]

https://github.com/oobabooga/text-generation-webui go to this github, and scroll all the way down

[gusty palm]

thanks i joined n asked

[nova remnant]

smol sus i was kinda hyped up about trying the new engine with SD but im getting the idea the extension could be buggy on some systems

[wispy canopy]

reinstalling venv helped alleviate that one

[lyric zealot]

oh is it?

[nova remnant]

btw ... all along til now i was rendering images at 768 then upscale it by 1.5x using highres, but now i decided to render at 512 instead and upscale by 2x... this made a remarkable difference how fast i can generate images

[wispy canopy]

@nova remnantyou got it working?

[lyric zealot]

I'm limited by ddrive, or just loading of profile is slow

[nova remnant]

no not yet, im still trying to find work around with tensor rt

[wispy canopy]

yeah, i got rid of the dll errors

[wispy canopy]

but now i i'm stuck on nvinfer.dll

[lyric zealot]

this is my speed on a 512x768

[nova remnant]

i dont have any errors showing up now, its that i can't seem to load the Tensor RT tab either

[lyric zealot]

maybe restart webui, and see if cmd have and errors?

[wispy canopy]

check the startup console

[wispy canopy]

you'll see that the tensor unit crashes

[nova remnant]

i did

no errors of that lkind i checked carefully

[lyric zealot]

[nova remnant]

my gawd smol sus my rendering speed is going like a lamborghini

[lyric zealot]

https://tenor.com/view/anime-speed-gif-20195114

[nova remnant]

im not kidding lol but yeah this new method made a huge differnce

[lyric zealot]

well, i might try it on my c drive (SSD) tomrrow

to see if that makes any difference

yup, legit huge

[nova remnant]

i have to be prompt now jkust to look at the preview images before it moves onto the next one

[lyric zealot]

now im just genning at 150 steps for the sake of it

[nova remnant]

and at this rate i could fill up a good chunk of my SSD pretty fast

[lyric zealot]

[nova remnant]

and uhh. this could not be some weird conincidence with tensor rt and driver upate, is it?

[lyric zealot]

is the generation faster with it even with hires fix?

[nova remnant]

yes almost by 3x

[lyric zealot]

no idea

okay

then it's definitely my HDD being stupid

[nova remnant]

3x no typo

[wispy canopy]

yeah, driver update was specific for the tensor rt extension too

[dawn jasper]

She is fabulous.

[nova remnant]

just the driver alone?

[lyric zealot]

for me, with hires fix, it will do this thingy

then it will be longer than normal generation without the new added tensor

[nova remnant]

i mean i dont have tensor rt working but would that driver make this change by itself?

[lyric zealot]

wait

im confused

you are using it with just the extension

but not with the driver?

[lyric zealot]

thanks as always

[nova remnant]

moi?

[lyric zealot]

[nova remnant]

my tensor rt didnt work with my webui, but i have the nvidia driver updated to latest

and im surprised how fast my image gen is going but using 512px and high res them by 2x

using lollypop too

[lyric zealot]

wait

then what changed in your case?

just driver update shouldn't have speed increase

[nova remnant]

only thing that i changed is enabling the super video resolution and setting its value to 2

that might have some impact on graphics apps

but idk

[lyric zealot]

[nova remnant]

actually let me try increasing the value to 4 max

[lyric zealot]

i don't think that matters for SD tho

not sure

[nova remnant]

yeah thats how i find it strangely intriguing

[lyric zealot]

well, imma head to sleep for now

[wispy canopy]

sleep well ❤️

[nova remnant]

sweet dream!

[lyric zealot]

[wispy canopy]

smoll sus, sdxl model?

[lyric zealot]

nope

all based sd1.5

[wispy canopy]

damn, that's some high qualilty then

❤️

[lyric zealot]

maybe cos some of them are generated with 150 steps

[wispy canopy]

and decent model 😛

[lyric zealot]

this is not tho v metadata (oops should point downwards)

[nova remnant]

wrong settings, i had old default upscaling of 1.5

[wispy canopy]

ok... got tensorrt tab

but building model crashes

OBVIOUSLY

[nova remnant]

damn

what so you suppose this is?

internal bug with engine?

[wispy canopy]

that you have increased speed, or that it doesn't work?

[nova remnant]

lol that it is buggy, in my case im assuming im facing extension bug

so this is what i notice when rendering at 512px, the face is gonna look coarse, contrary to rendering at 768px

[nova remnant]

@nova remnant

I slept now people are tensor rt'ing

Did you guys get it to work

[nova remnant]

no luck on my end :/

cause i couldn't get the extension to work

@supple raptor did you see about that tensor rt

nubby is the expert

[nova remnant]

i should stop but so many cute pics

[supple raptor]

yeah, people were using tensor rt a while ago. the speed seems real nice, but it's super inconvenient. I have no interest in compiling a new model each time I want to use a different lora or change the resolution.

[supple raptor]

it would be great if you wanted to run a discord image generation bot or something like that though.

[nova remnant]

nubby can you explain what you mean by having to create new model everytime ?

[nova remnant]

Legs

[nova remnant]

this came out better

In today’s Game Ready Driver, we’ve added TensorRT acceleration for Stable Diffusion Web UI, which boosts GeForce RTX performance by up to 2X. This enhancement makes generating AI images faster than ever before, giving users the ability to iterate and save time.

[nova remnant]

today -> 17th

[supple raptor]

tensor rt is converting models to .onnx when in this format the model has a min, max and optimal batch size, resolution, and prompt size. loras also cannot be added on the fly so if you want to change the weight of a lora or change it to another one, you have to completely recompile the model.

[nova remnant]

that would be crazy to use tensor rt then

a sd 1.5model over 2gig and sdxl over 6gig

[supple raptor]

the people who write release notes often don't really know what they're talking about. I'm sure they did something related to tensor rt in the new driver, but it has been possible for a while.

[nova remnant]

idk if i should bother with tensor rt unless there is something that's not clear in their writings

[supple raptor]

yep, this is why I don't have much interest in actually running it. it's cool, but the downsides are too big.

it also breaks a lot of extensions afaik

[nova remnant]

i was excited about having to render at 2x speed

[nova remnant]

If you just type a new prompt, do you have to recompile again?

any thoughts on the newest driver today if that has any impact on sd ?

[supple raptor]

no impact if not using tensor rt.

[oak spoke]

huh didn't expected this

[supple raptor]

no, but in some cases you should for different lengths of prompts.

[oak spoke]

the download to like ratio is crazy

3.9k download

that's almost 1 in 3 download is a like

[nova remnant]

Hello juice man

[oak spoke]

hello catto

[bitter latch]

tried the tensorRT again got

[nova remnant]

i gave up too

[bitter latch]

closed and reloading

waits

[nova remnant]

it almost broke my webui

[bitter latch]

yep got the entrypoint not found error.. again

[nova remnant]

:/

i had to delete the venv folder to fix the crap on my end

[supple raptor]

do you have an rtx gpu @bitter latch?

[bitter latch]

RTX 3050

hmm well uninstalled that crap now working on error GradioUnusedKwargWarning

[supple raptor]

ok cool, it should be able to be used then. I haven't bothered to install it so I wouldn't be much help in troubleshooting it though.

[bitter latch]

oh ok this is a dev reminder error. doh

[bitter latch]

so from what i'm gathering it's not big then, whew

[supple raptor]

it's a big increase in performance, it just comes with a lot of it's own downsides

[nova remnant]

nubby any reason in general why everyone should not render images at 512px on sd 1.5 then perhaps high res fix by 2x and further upscale if needed? cause im getting a significant speed boost and time reduction on render with 512px gen

[supple raptor]

I'd rather wait longer for each gen, but be able to change whatever I want whenever I want.

[supple raptor]

I don't personally like 1:1 aspect ratio images typically, but other than that no, this is a fine workflow

I do 544x960 because at 2x highres fix it leaves me with 1088x1920 images so I only have to crop 8 pixels of the small side to make it 1080p.

[nova remnant]

i often switch between 512 640 800 960 as base standards before high res

like this is 512x640

[nova remnant]

SDXL can't replace SD 1.5

lot of factors involved in that

from technical usability aspects to even aesthetics

[supple raptor]

aesthetics are just down to needing finetunes with the right aesthetics to be made. it's too demanding hardware wise to get the same level of use that SD1.5 gets though.

[nova remnant]

yes, that's the thing I feel about technical usability

[nova remnant]

is that miku?

[bitter latch]

changed from invokeAI back to xformers. seems to work better

yeah

using a cartoon real merg

[nova remnant]

hatsune miku or nakano miku?

[bitter latch]

shrugs. just put Miku

[nova remnant]

haha

[supple raptor]

yeah, invokeAI memory opimization isn't good. if they are usable on your card you should be using either xformers or SDP and both are available for you.

[nova remnant]

she is hatsune miku

[bitter latch]

switches to AOM3_orangemixs model

[nova remnant]

i think she is angry cause of the fingers