🍥｜anime

[nova remnant]

brace yourself

[steady grail]

ok very appealing

[nova remnant]

im gonna swap model and see if this problem still there

[steady grail]

is adding meaningless details

maybe you need a different face model for xl models

[stable crane]

happy friday~

[supple raptor]

@nova remnant That looks like the resolution is too low on an XL model. Try specifying a resolution for the inpaint that's closer to 1024x1024

[nova remnant]

im using 768x1024

[stable crane]

very nice

new model @supple raptor ?

[steady grail]

very d&d

[nova remnant]

its the same thing with other model too

adetailer when on it messes up the face

[supple raptor]

That's Ao's model https://civitai.com/models/211628/soushiki

[steady grail]

i love that model too

[nova remnant]

i dunno if i should be happy or unhappy to get good faces w/o using adetailer

[steady grail]

just become one of my favorite

[nova remnant]

should this section matter, like should i increase them?

[steady grail]

looks like the face gets hooked well enough

[nova remnant]

yes from the marker i see it does

[steady grail]

it's just that the model is not doing his job

[supple raptor]

Set them to somewhere close to 1024x1024 and click the checkbox right above them.

[nova remnant]

that made a slight improvement .. but compared to what i get before adetailer modifies, something is visibly off

[astral jasper]

alright bye

[stable crane]

@supple raptor do you know what's up with the VAE it recos?

[supple raptor]

It's just the standard kl-f8-anime2 vae from WD. I use it with my blessed down version that I use on everything.

[stable crane]

ah ok

[nova remnant]

if a model has baked VAE would that matter if i selected sdxl vae from a111 menu?

[supple raptor]

selecting a vae in the dropdown just overrides the one that's baked in

[nova remnant]

so i can then set to Automatic instead of sdxl vae?

[supple raptor]

I'd recommend setting none if you want to use the baked in VAE. Automatic is intended to load a VAE that's in the model folder with the same filename as the model. It generally will correctly use the baked in VAE if it doesn't find one that shares the same name, but sometimes it will bug out and just reuse the last one that was loaded.

[nova remnant]

wow.. ok setting that to none now

thanks

[supple raptor]

🍝 code

[steady grail]

gonna go to sleep

bye

[terse ore]

Made this from a blender screenshot by importing characters from a game and used controlnet+img2img to make it look more 2d. Color is not even and it's far from perfect

[supple raptor]

I finally ordered a drawing tablet. Soon I won't have to fix my images with mouse

[nova remnant]

i wanted to get a tablet for drawing too, only thing that kept me is that im bad at sktching with hands

[supple raptor]

I used to do it a little, just haven't for a long time. Maybe I'll get back into it once I get comfortable with the tablet.

[nova remnant]

that'd be good for others who will use your fine tuned models... as for me its -not- an option to use drawing tablet, i have no concept of keeping proportion in drawing especially with human anatomy

im kinda that bad as sd is with fingers

[terse ore]

I have a cheap drawing tablet, but I used to use the old Samsung android tablet. It wasn't too bad. It can be used to fix AI stuff I guess

[nova remnant]

nice ... but i don't think it'll work for me lol

[terse ore]

Yeah I'm bad at it too

[supple raptor]

The only way to learn to draw is to do it.

[nova remnant]

true, but im kinda augmented with SD

[supple raptor]

That just means you can draw something, then run it through img2img/controlnet to refine it

[nova remnant]

well tbh im very visual :p

but not too savvy on the technical application that requires manual interactions

[nova remnant]

https://huggingface.co/bluepen5805/blue_pencil-XL 👀 the author merged all the XL Models in the whole Civitai database

[stable crane]

how do you do a temp merge in a1111 again?

[nova remnant]

no clue about temp merge, i only see checkpoint merger there

i think im going in circle with merging models :/

[native halo]

https://github.com/wkpark/sd-webui-model-mixer this

[nova remnant]

💤

[terse ore]

Is sdxl img2img+controlnet better than 1.5?

[native halo]

better at what?

[nova remnant]

not a tantrum, she is waking up from sleep, yawning and stretching

[supple raptor]

SDXL controlnet is worse than 1.x controlnet from what I've heard. I haven't tried it myself or looked much into it.

[native halo]

its better at eating your vram,(it likes to eat more at a1111)

[supple raptor]

The controlnet part shouldn't since it's all in LoRA format

[native halo]

there are controlnets for xl that arent loras like ipadapter that eats your ram because auto1111 optimization sucks for xl but if u have enough u prob wont notice

on comfy is better if u like to play with the noodles

[nova remnant]

there are t2i models for SDXL controlnet .. lot smaller in size and takes lot less resource

each of those models are like 150ish mb

https://huggingface.co/lllyasviel/sd_control_collection/tree/main

[native halo]

now run 3 models at the same time ip adapter+openpose+depth

[terse ore]

Ah, I see. Because sometimes controlnet messes up fingers so I thought maybe sdxl would be more accurate at that sort of thing

[native halo]

yea xl is better at that u get less deformities

but its better if u use comfy for xl,its more optimized

[nova remnant]

actually let me benchmark the basics with t2i models using canny and depth

[native halo]

ipadapter was the one that was eatin more vram when i tried 3 controlnets,try with that one too

[nova remnant]

canny and depth

well t2i is the tech that surpassed other types of models

you should use t2i for efficiency

[native halo]

no,ip adapter is better for mixing images

[nova remnant]

this one took me 28s

if ip is taking lot more time i dont think that's better optimized

and i didnt use lcm lora this time

[native halo]

it takes more time but its better for mixing images like u have an image of an elephant and another img of a girl and u want to mix both into a single img

or if u want to add something that sd cant do

[nova remnant]

which do you use for that example?

i mean the control type

gimme the secret lol im learning

[native halo]

ipadapter https://huggingface.co/lllyasviel/sd_control_collection/blob/main/ip-adapter_xl.pth idk if theres other ver that more optimized

[nova remnant]

no i mean what are the control type do you use for that girl and elephant example

[native halo]

control-type ip-adapter

u have to use 2 models in a1111

[nova remnant]

ok let me check

[native halo]

in comfy is easier

[nova remnant]

i think i need to see guide on the process for merging two different characters into one

[native halo]

yea on comfy is better on a1111 idk how it works

on a1111 u can use one model to add something to the img that u cant get tru normal means

[nova remnant]

i didnt do my homework on controlne extensively

which is why cross mixing two different characters into one is unfamiliar to me

and i understand this is not quite using roop or reactor just to swap faces

[native halo]

here is a guide on how to do it in comfy https://civitai.com/articles/3099/combine-multiple-images-with-ipadapter-a-workflow-for-comfyui

idk about a1111

prob cant be done there

what can be done on a1111 with ipadater and i tried it is adding something to an image tru ipadapter

[nova remnant]

im not sure i understood the concept ... but let me try to explain what i understood, i load up two different images in controlnet to merge them together?

and none of the videos on youtube about ipadapter were much help

[native halo]

yes but on auto1111 u cant do that only on comfy

what u can do on a1111 is load an img on controlnet enable ipadapter and it adds the thing on the img to the output in text2img

[nova remnant]

so, i load up an image in contolnet, select ipadapter and add stuff on the image with text prompt?

can't i just use two instances of ipadpater ?

a 3 gig file is downloading right now so i have to wait for a while before i can put that to test

[native halo]

yea u can but on a1111 u cant use 2 instances of ipadater to combine 2 images into one

[nova remnant]

oooooh.. thats not good

[native halo]

but u can use it for example u prompt "a girl sleeping on a red couch" then enable ipadapter and put there the img of a red sofa then it will add that same sofa from the img to the output img

[nova remnant]

ok here is the fun thing i did that i showed you the image for ....

original image loaded on canny and depth ... then used the text prompt
pretty female cyberpunk hacker with blue colored glasses in a jacket with a Beanie, long grey hair, half t-shirt, ripped jeans, (poster art:1.3)
and got that image i posted above

but that was using t2i

[native halo]

try to prompt with an anime XL model, a girl lying on a red sopha then enable controlnet > ip-adapter then add this img to controlnet to see if it adds the sofa to the output img

[nova remnant]

ok

[ornate echo]

sopha

Try "sofa".

[nova remnant]

a coarse result, i was experimenting and obviously the method wasnt proper but here is the output

[native halo]

sophia

[nova remnant]

i changed that to couch haha

[native halo]

i got this one but reduced weight of controlnet

[nova remnant]

im having inteference with that model download in console

and its locking up on a1111 controlnet

almost done, i'll try again

that was the ip model for xl

[native halo]

yea it probably eat lots of ram/vram on xl

[nova remnant]

got distracted, had an earthquake

both t2i and ip control types put the subject in couch but im not getting quality results

[native halo]

try to reduce the weight

[nova remnant]

reduced to 0.6 but weight not be the issue, the text input that generates characters is coming out poorly

[native halo]

you could try to enable a 2nd instance of controlnet and use the img i posted as an openpose input

[nova remnant]

its downloading another file

but almost done

what im going to do is load up the couch on one instance of ip-adapter and load up an image of a girl in lying position on another instance of ip-adapter

https://youtu.be/KHm5Q5TfNvE

the example he is showing i can do that already w/o ip adapters

just using canny and depth

what i'm curious about is taking two different images and merging them together meaningfully

[keen hollow]

evenin'

quiet...

[native halo]

https://tenor.com/view/family-guy-peter-griffin-happy-dance-gif-11863866965971130432

[keen hollow]

what's new @native halo ?

any new steven segal images being generated? 😛

[native halo]

#🍥｜anime message yes this one 🦶

[keen hollow]

that is a foot

[nova remnant]

https://youtu.be/oWZblDTfsQw

[nova remnant]

sdxl doesnt have face model CN for ip adapter

this is going over my head https://youtu.be/yrqT5FaBoIU

[keen hollow]

people chatting in here is greek to me now...

feels like everyone moved to comfy and talks about IP Adapters and ControlLMGizmos and whatever else.

[nova remnant]

im never switching to comfy :p

[keen hollow]

isn't all the new turbo/live stuff only in comfy?

[nova remnant]

no

you can use that on a1111 too

[nova remnant]

but truth be told .. comfy is lot faster to implement new features as they come, a1111 dev team are lazy af

[keen hollow]

I don't think auto has updated in quite some time

comfy, to me, just feels like I have to do 8 things that take 1 step in auto

[nova remnant]

their last update iirc was few weeks after sdxl was launched

[nova remnant]

i cant follow where the wires go

[keen hollow]

yeah it gets wild

[nova remnant]

lol

[native halo]

my try

[nova remnant]

nice, what did you do?

[native halo]

ipadapter + openpose

[nova remnant]

ahhh

so you used 2 images?

[native halo]

yea the other i posted earlier for the pose + the couch pic

[nova remnant]

nice

i got discouraged when the character came out badly

so i stopped

[native halo]

yea u have to play with weights until u get a perfect pic

[keen hollow]

ok, what is ip adapter 😐

is it a controlnet model ?

[nova remnant]

yeah controlnet type

[keen hollow]

haven't messed with that

[nova remnant]

me neither, just trying it out now after steven mentioned it

when you enable the hires fix and upscale by 1 .. meaning no changes in image size, would that still add some detail to the image by denoising strength of 0.4 ?

[native halo]

idk havent tried that

[nova remnant]

i see several youtubers using that method

[keen hollow]

try it and report back

[nova remnant]

lots and lots of deformity

[keen hollow]

well you gotta share

we wanna see the hands

[nova remnant]

nawww they are hidous defect

hedious

wow i need sleep

hideous

but soon as i turn off hires fix the deformities go away

[keen hollow]

try lower denoise

[nova remnant]

i kept that to 0.3 and 0.4

not using lcm lora

a woman with flowy blonde hair in the garden watering the plants, she is wearing cute summer dress. the sun is shining with puffy white cloud.

[keen hollow]

ill try it in dalle

[nova remnant]

the context is cool

the character is actually focusing on planing the water

.. wtf.. watering the plants

i gotta get sleep 😐

[keen hollow]

me too

but not yet 😛

[nova remnant]

my brain is kinda fried

was merging models for hours

ironically nothing came out good

[nova remnant]

while experimenting with models i bumped into something accidental ....

[steady grail]

fantastic

this is dal-e?

morning

this is so berautiful, very abstract, but beautiful

[nova remnant]

its pretty neat how dalle delivers context

[stable crane]

dalle always delivers

[native halo]

try to make a tiny girl standing next to a giant white cat statue on XL

[native halo]

now change statue to one of maneki-neko

[nova remnant]

i dunno what maneki neko is

and look this is better

[native halo]

its this one in japan

[nova remnant]

lol i dont think sd will recognze that

[native halo]

maybe with ipadapter

[nova remnant]

a chibi girl standing next to a statue of a giant maneki-neko

[gusty palm]

wat model is that

looks coo.l like irl plastic figurines

[nova remnant]

lol just some radnom xl model let me get you the link

https://civitai.com/models/125907/realcartoon-xl?modelVersionId=234137

i was also trying out blue pencil ...has most of the xl model from civitai merged into it .. but didnt have one i linked above

[native halo]

this one was with controlnet

[nova remnant]

looks awesome

[stable crane]

cute

[nova remnant]

you can touch the cuteness in that

[ornate echo]

disembodied hand

[native halo]

yea too low weight

[ornate echo]

😢

[gusty palm]

cool

[gusty palm]

i wanna try the 1.5 version cuz i only have like one of those and can never remember which one it is

[gusty palm]

i already have them aha i needa use those more

[gusty palm]

wonder if it would merge well into nmy current merge

[nova remnant]

katie i lost sleep doing that lol

[native halo]

theres one like that for 1.5 that looks like plastic but i forgor name

[supple raptor]

a1111 dev deam has just been working on the dev branch. last commit was 29 minutes ago and it's 399 commits ahead of master. Whenever the next update lands it should be a pretty big one. I haven't really been following along with everything in the dev branch though.

[nova remnant]

holly ..... that sounds big

im glad to hear that 🙂

a (siamese cat:1.1) chasing a tyrannosaurus rex across the valley.

[supple raptor]

that cat is definitely not chasing the trex XD

[native halo]

make the cat giant and the t-rex small

[nova remnant]

lol the cat is siamese .. pretty fierce

siamese cats grow pretty big too 🙂

a girl at a park, roller blading and having fun

any chance they might work on LCM LoRA any time soon?

these images im generating are not using lcm lora btw

[lyric zealot]

@dull jackal @finite herald #🍥｜anime message

[nova remnant]

i discovered i can generate crispier image the regular way w/o lcm lora

[steady grail]

@dull jackal a scam

useless to say, don't follow what that link says

even better, don't open it

[lyric zealot]

it's kinda sad that the user likely got his account compromised by maybe clicking on links like this

[steady grail]

yeah

i had a friend in that exact situation

i had to report his account to steam, to limit the damage, not that they did something

[lyric zealot]

same

[native halo]

if i had steam i could lose my entire acc

[steady grail]

monthly change the pass of your mail and account

[native halo]

https://tenor.com/view/blue-archive-takanashi-hoshino-hoshino-soccer-football-gif-17053111496513313608

[steady grail]

tha phrase is a parody, or really means something?

[nova remnant]

i can't make the pup follow her from behind

[nova remnant]

lol whats this

[steady grail]

layered images are the most hard to do

[native halo]

🙏 🙇‍♂️

[steady grail]

what alhamdulillah means?

[native halo]

praise to god or thank god

[steady grail]

oh, so it's not a joke, it's a real word, sorry for my ignorance

what language is that? his use is becoming more common and i would like to deepen my knowledge

[native halo]

arabic

[steady grail]

arabic is a common language? or there is a more common dialect?

[native halo]

it divides in other 4 languages

[steady grail]

as expected

i love to deepen my knowledge when i'm relaxing before starting the day

25 min and starts my day

[nova remnant]

btw about hires fix upscaled by 1 it does have impact on image due to its steps and denoise, i had 25 sampling steps and hires 10 steps

[steady grail]

like an img2img i suppose

[nova remnant]

not enlarging but adding detail i think

[steady grail]

takes the image, adds noise and the output is a mix of the 2

[nova remnant]

ahh, but i havent spotted a visible difference in that method

[steady grail]

makes sense in a latent upscale, but in a value at 1, is just lightly different output

at least, i think so, take it with a grain of salt

maybe there are steps i ignore

[nova remnant]

lol im not sure about it... if that actually works by adding noise to make images better and not hogging resources by upscaling then its good for my case

[steady grail]

i suppose so

but i think there are better ways to spend that time in improving the image

[nova remnant]

hires upscale is good but i can't use that when running sdxl

[steady grail]

auto1111 or comfy?

[nova remnant]

auto

[steady grail]

on auto, i suppose a pass on img2img with different loras at low denoise can give better results

[nova remnant]

heard from nubby auto dev are working on releasing a huge update

[steady grail]

good

i hope they push away that dogma of nohalf fix

[nova remnant]

which one is that?

[steady grail]

i moved on comfy for that reason, that reason

with my gpu, i had this black gens bug

[nova remnant]

oh i heard about that i even had that issue

[steady grail]

they fixed it, but the fix they used, absorb too much resources

so, i moved on comfy

[nova remnant]

are you talking about --no-half-vae?

[steady grail]

yep

--no-half --no-half-vae

for a certain version, xformers fixed that problem, without using that solution

and that reduced the resource consumption by a lot

[steady grail]

but with the successive updates, the thing broke again

[nova remnant]

:/

lets hope they sort that out in the upcoming update

[steady grail]

i holpe so, but i doubt, that's why i called it a dogma

with comfy i never had this problem

they faced that problem differently

i don't know how, i can frely generate images at 2024x2024 freely

[nova remnant]

a cottage inside a soup bowl

[nova remnant]

how long does that to take generate 2048 x?

[supple raptor]

yep. although, probably a bit more effective because it's still upscaling to whatever the native upscale factor of the upscaler is (likely 4x) then downscaling back to the original dimensions which adds noise and that noise sort of gives the denoising something to bite on and create additional details.

[steady grail]

don't remember

i try a 1024x1024

it should take 15-30 min

[nova remnant]

15 mins for 1024?

[steady grail]

if i don't remember wrong

i try to gen something

i have a 4gvra

[nova remnant]

ok

[steady grail]

i a very low end gpu for this business

[nova remnant]

when i used 4gb i had to stick to 768x max

and it took me near about 5 mins for each image

[steady grail]

i'll look for a cool prompt and try to gen something

[slate narwhal]

lovely colors

[steady grail]

upscaling and downscaling? like kohya?

[supple raptor]

When you see the 2x or 4x in the name of an ESRGAN upscaler, that is telling you the factor that it will upscale by. There's no way around it, it just does that. The way Auto achieves upscaling by any factor is by running the upscaler then resizing the resulting image to match the upscaling factor that was defined by the user. If you use Comfy, you have to do it the same way with first an upscaling node, then a resizing node.

[steady grail]

oh

[steady grail]

looks like very time consuming

i should try it, it may give interesting results

💡

soushiki is really mind blowing

i really love this model

[nova remnant]

nubby the glitch i had earlier with extra noise on character's faces was due to high weight on lcm lora

[steady grail]

ohhhh

[nova remnant]

and the samplers don't all treat the lcm lora weight at same value

[steady grail]

using too much lora on adetailer usually is a bad idea

[nova remnant]

no not the adetailer .. it was the lcm lora weight

lora:lcm-lora-sdxl:0.3

[nova remnant]

i had it up to 0.7

[steady grail]

weird faces without adetailer?

[nova remnant]

i was also messing with ohter settings

right now im using lcm adetailer but with lower lora weight

and faces come out smooth

also the same checkpoint i was trying earlier

[nova remnant]

what you mean you can not draw with a mouse

thats unethical

also hewo chatters

[nova remnant]

plus the hires too now

[steady grail]

i first draw on paper and than use the drawing board

[supple raptor]

What else am I supposed to do with them? They aren't getting my cheese without working for it.

[nova remnant]

https://tenor.com/view/cheesed-to-meet-you-gif-21409625

[steady grail]

i honestly use adetailer wiiht only style lora and add details at low weight

[steady grail]

1.376x1.376 in 25 min

on a 4gvram

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1180412257239318648/00044-1249798613.png?ex=657d5395&is=656ade95&hm=d1faa174fd2a6d68175ab5b2b0562a907266c8e12cdb8c00a54fc97b04e31ce2&=&format=webp&quality=lossless&width=441&height=662

[steady grail]

loading in lowvram mode 1701.657901763916
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [01:06<00:00, 3.34s/it]
Requested to load AutoencoderKL
Loading 1 new model
Prompt executed in 1624.82 seconds

[steady grail]

688 should be the max i can do, with my vram, without gettting impratical slowdowns

oh right, loras, slowed down the gen

without loras, is even faster

oh damn, how much i love this model

it blows my mind how cool the details are

[vital raptor]

I can't imagine trying to do this stuff on a 4gb card, I get twitchy when an image takes over ~15 seconds

[steady grail]

eheheh

i only upscale images i really like

[gusty palm]

which ai did u ask about the merge model forumla yesterday?>

that explanation was helpful but didnt seem totally accurate for the model merger addon

[vital raptor]

my default generation on SD1.5 is 1024x1536, takes 4s

I upscale from there if I like it

[steady grail]

lol

[steady grail]

such raw power costs too much for me XD

[vital raptor]

yeah, it was expensive

[steady grail]

anyway i'm more in studying new ways to generate

that's why i like comfy so much

every day i test new workflows, without generating anything really meaningful

[vital raptor]

I don't generate anything meaningful most days, spent the last couple days trying to learn how to train SDXL properly

[gusty palm]

i love how this merge comes out, im still not commited to adding the loras into it thogu which are doing heavy lifting.
muskrat sisters and ground hog girl

i wanna try adding in that 3d pixar model and see what it does

[steady grail]

a little baked

[vital raptor]

merging and mixing models is fun too

[gusty palm]

9 cfg and using a minimal vector style lora

ill lower its weight and try 7

[steady grail]

9 should still be inside the tollerance limit of a model

pretty sure some fine tuning with weights will improve the model

till now, it looks good

[gusty palm]

https://civitai.com/models/112053
this too adds alot

https://civitai.com/models/103904

using bofam

[gusty palm]

tiger fox

i like the previous gen actually bettert

[gusty palm]

silver fox

[gusty palm]

upscale with CADS on vs off

o wow i couldnt see the difference if i hadnt switched between them

redoing that without the loras in the prompt

[supple raptor]

cute!

[vital raptor]

silver foxes are nice 😛

[gusty palm]

aaw its like the same girl in a different universe

https://i.imgur.com/5vYjXUk.gif

CADS produces weird results on the initial gen but i like it for upscaling now

some of them are not bad though

[vital raptor]

looks a little burned, but the style is cute

[gusty palm]

ya i dont quite understand how the settings work

i like what its doing im img2img though

[steady grail]

cads? what's that?

[gusty palm]

its some type of noise modifier or something >< im not super sure, it seems similar ot CD tuner but idrk

[supple raptor]

damn, i was going to read the paper to try to understand CADS but it's 33 pages long. I'm not that interested. Here it is if anyone else wants to look it over. https://arxiv.org/abs/2310.17347

[toxic dragon]

Comparison between the two styles I'm between rn.

[nova remnant]

I'm assuming the darker one has detailer/more details in it ?

[nova remnant]

Lmao is that a lora or did you just prompt for a card

[jagged wasp]

lol just anime driver license

[nova remnant]

X3 there's always something I'd never think about for SD to generate

[gusty palm]

oh wow

[toxic dragon]

yes

Lookin to see which people think is better

Left or Right

[steady grail]

on the right we have better details, but a weird darker face, on the left, the color are to high in contrast

oh, cads is a sampler

[supple raptor]

I'd say left, but I would run an inpaint at full res on the girl in the background since she's lacking detail. I think somewhere between the two would be best.

[steady grail]

i didn't generated decently from so much time

generating images looks almost like a miracle

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1180447978469740664/00119-1775441710.png?ex=657d74d9&is=656affd9&hm=7d3673c51ac49af7410c40f6d45aa603965401bb6874560d6475b4db27fe9d0a&=&format=webp&quality=lossless&width=441&height=662

[toxic dragon]

Yeah that's probably what I was thinking to do later

Since I haven't really inpainted much until recently.

Found limited success with it

[steady grail]

zooming in high res images is like looking for woldo

noticing all the little details added by the model

[toxic dragon]

True.

[supple raptor]

If you're using A1111 or SD.Next this is the best inpainting experience https://github.com/poipoi300/auto-sd-krita-ext-xl.git

If you're using Comfy then good luck because Comfy inpainting is cursed AF.

[steady grail]

it's magic and relaxing

[toxic dragon]

Speaking of waldo.

I remove them and there's a ghost in the forest.

The demon is gone.

[steady grail]

and the small bird next on the right?

XD

[bitter latch]

she has been summoned

[toxic dragon]

?

[bitter latch]

sets a large batch going with some tweaks and goes back to bed.. thud

[steady grail]

on the tree on the right and a little leaf on the hair lol

[toxic dragon]

You mean that tiny green speck below the ribbon?

[supple raptor]

what bird?

[steady grail]

bird's nest

[toxic dragon]

https://tenor.com/view/spy-kids-lemme-zoom-in-on-that-zoom-watching-you-spy-gif-5212510

[toxic dragon]

Let the birds chill bro they aren't bothering anyone.

[steady grail]

it's just fashinating

pretty sure is a glitch between the tree and the hair lol

[supple raptor]

I'd maybe fix the green spot in the hair, the brown patches of leaves aren't anything out of the ordinary though.

[nova remnant]

https://tenor.com/view/enhance-photoshop-gif-4844535

[toxic dragon]

I'm more concerned about the weird hair on over the straps

The hair doesn't look natural over the shoulders and straps

this is a little better.

[nova remnant]

if that adetailer is at 1 strength, I'd put it at 0.2 I like some or without, its both good

[steady grail]

the difference of just 20 pixel lol, damn

[nova remnant]

How many steps is that gen

[toxic dragon]

20 I think?

10 hires

[nova remnant]

what sampler ?

[supple raptor]

yeah that's pretty good, i'd probably be happy with it if it were mine. you could paint over the top strand of hair where it crosses the strap on the right since it's slightly messy but that's getting pretty nitpicky.

[nova remnant]

maybe try more if its not like euler a which doesn't do a lot of steps thing

[toxic dragon]

I'm using euler a rn

[native halo]

needs smaller hair ribbons like these

[supple raptor]

euler a likes roughly 30-35 low res and 20ish hires. it needs a bit more than the more modern samplers like all of the DPM++ variants.

[supple raptor]

last change i'd make is mirroring the image because the logo is too close to the adidas logo but it's backwards.

[nova remnant]

lmao

[nova remnant]

flaps

[toxic dragon]

Yeah but I want the ribbon on the right side of the image.

[supple raptor]

fair. the logo isn't really a problem in that case

[toxic dragon]

decided to clean up the face a little instead and hair strands a bit.

the rest seems to be superficial

the left side hair falls well enough on the strap that its tolerable.

[nova remnant]

[supple raptor]

Yeah, I don't see any real problems with the hair on the left strap.

[native halo]

my honest reaction

[toxic dragon]

I feel like I'm finally getting progress despite still not using much of my new card

Although its really hard to find something in or near the style I'm using to where my best bet is to maybe consider training something

[toxic dragon]

One more higher detailer attempt

[gusty palm]

its crazy just how many variations of the same image can be made

[steady grail]

guys, after the update of comfuy i get this error when i generate
:\Git\comfyui clean\ComfyUI_windows_portable\python_embeded\lib\site-packages\torchsde_brownian\brownian_interval.py:594: UserWarning: Should have tb<=t1 but got tb=14.614643096923828 and t1=14.614643.
warnings.warn(f"Should have {tb_name}<=t1 but got {tb_name}={tb} and t1={self._end}.")

someone know something?

[supple raptor]

it's not bad, but I still prefer the previous version.

[supple raptor]

I think it shouldn't be a problem since it's only a warning, but if you're getting unexpected results then you should open an issue in the comfy repository so he can address the issue.
If you are getting unexpected results then it would probably be worth looking through the open issues for any custom nodes you're using first to see if it has been reported in any of them.

I've never used SDE optimization in Comfy so I have no idea if it's common. I use xformers.

[steady grail]

no... at sight, the warning looks harmles

sde?

oh, sampler

[supple raptor]

TorchSDE is a memory optimization that is built into pytorch. It is very similar to xformers but is a bit less memory efficient. If xformers is present in the venv you're using for comfy it should default to xformers from my experience, though you should be able to force it with the --xformers launch arg.

[gusty palm]

does the 'make it more' thing work with sd? or is there some way to approximate it

[steady grail]

let's try it out

[gusty palm]

seems like ip2p

[supple raptor]

sort of... that originates from InstuctPix2Pix models which are supported by A1111, but they aren't very good at anime in my experience.

[supple raptor]

yep, exactly

[supple raptor]

do you have A1111 installed?

[steady grail]

a very outdated version

[gusty palm]

yeah they give no anime examples hmm

[steady grail]

why?

[supple raptor]

ah, then you probably won't want to use it's venv. I was going to send a .bat file that would launch comfy with your A1111 venv which would have presumably already had xformers installed.

[steady grail]

i moved to comfy exactly do not have to deal with auto any more XD

paradoxically that version of xformer on the outdated auto should be the most compatible with my gpu

[supple raptor]

tbh, A1111 has been the least headache for me and I have A1111, SD.Next, and Comfy all installed.

[steady grail]

auto is not very gentle with my 4gvram

[supple raptor]

Well, if you want to use it here is the .bat
You'll need to edit line 7 to match the path to the specified file in the venv that you want to use. It would also work with a python venv that's in comfy.

[steady grail]

i can try, just for curiosity

[supple raptor]

make sure you place the .bat file in the root directory of your comfyui install. the same folder that has your main.py file.

[steady grail]

i think the launced a fast update and one error vanished, but now i have this XDmissing {'cond_stage_model.clip_l.text_projection', 'cond_stage_model.clip_l.logit_scale'}
left over keys: dict_keys(['cond_stage_model.clip_l.transformer.text_model.embeddings.position_ids'])

i think it's because i don't have transformer accelerator

[supple raptor]

that looks like an outdated transformers library which is not the same thing as xformers.

[steady grail]

but the last time i looked in it, transformers was only for linux

[supple raptor]

which gpu do you have?

[steady grail]

i don't know if someone complied a transformer for win later

nvidia geforce gtx 1650

[supple raptor]

oof. the lack of proper fp16 support really hurts.

[steady grail]

yeah i know, it's a miracle i can play with ai using this gpu

i had to finetune comfy very much

[supple raptor]

it's kinda sad because power wise the 16xx series isn't too bad, but then it can't do fp16 so it has to work twice as hard.

[steady grail]

yeah, true

i had to do many work around

to improve performance

[steady grail]

but after all, i can say it's funny

[supple raptor]

a 50 series still isn't going to be amazing, but it would be a lot faster at fp16 than at fp32

[steady grail]

mh...

[gusty palm]

id like something like this . i tried comfy but couldnt figure it out but wanna try it again

or actually idk i had it working before so idk if id need it

[supple raptor]

the .bat file mainly makes it easier to launch since you just have to double click the .bat like you do with a1111's webui-user.bat and avoids needing to have another 3+gb python venv on your drive. You're more than welcome to use it. I'm not that well versed in actually using Comfy, but I can help you at least get it running with that .bat if you want. It also has a place for your launch args just like the webui-user.bat if you want/need to specify any.

I know it well enough to make a txt2img or img2img workflow with or without hires fix and refiner, but things like adetailer and controlnet are beyond what I've taken the time to figure out.

[steady grail]

if i did knew i would fall in love with ai, i would have looked for a better gpu

[supple raptor]

I upgraded from a 2070 to 3080 12GB because of AI. I was happy with the gaming performance I was getting but wanted more vram and speed for AI.

[steady grail]

eheheh

having the money, i would have done the same thing withou a doubt

i will do the same thing without doubt, if they stop taxing all my savings >.<

[supple raptor]

If I could go back in time I would have got a used 3090 for $100 more instead. I thought 12GB would be plenty, but I can easily run out.

[nova remnant]

[supple raptor]

poor little smol vram frens

[steady grail]

eheheh

now that i think about it, i'm saving for a new piano....

fuuuuuuuuuuuuuuuuuuuuu....

why all my loves are so expensive?

[supple raptor]

oof. yeah any serious music equiment is expensive af

[steady grail]

they are....

expecially when you fall in love with quality

every penny is worth, on a musical instrument

[supple raptor]

yeah, I mean even a budget piano isn't that cheap. hell a mediocre keyboard can be fairly expensive.

[gusty palm]

before ai i was always slightly upset at the amount of seemingly unused resources GPUs carried in regard to gaming

semeed totally overkill to have 6+gb when ive never seen a game come close to using anything close to that

[supple raptor]

I'd probably look at estate sales if I was in the market for a piano. Old people like pianos and it isn't too uncommon to find them left to a family that has no interest in them.

[steady grail]

my first piano was 500€ and i was about to buy a 700€ one, but every one they begged me to try a cheaper one first, even if they were my money

[gusty palm]

u can get a 61key midi keyboard for less than 100$

[steady grail]

not worth it

[gusty palm]

how not

[steady grail]

even the seller begged me

the feeling on the keys is different

[gusty palm]

ah

[steady grail]

very different, and the nuances of the sound is totally different

[gusty palm]

yeah they have weighted keys versions but theyre not under 100$

[steady grail]

the most cheap is 300€ and they are crap

honestly

the most cheap good piano is min 500€, the mine

[gusty palm]

i love my devices

[steady grail]

the 700€ was a mid top tier

what a good toy

i loved the 700€ one, but when i heard the seller to try something cheaper, as a noob, i accepted

still don't know if it was a good choice

but anyway, i still love that instrument

it gives me feelings

oh god

[supple raptor]

I tried guitar at like 5yo and quickly realized I have no musical talent. I feel I dodged a very expensive bullet

[steady grail]

uahaha

[supple raptor]

I'm content with singing in the shower

[steady grail]

musing is something of natural, or you are born with it, or you are not

[gusty palm]

nah its just practice

guitar is hard my fingers dont like it

[supple raptor]

i love music, but I can't identify a note to save my life.

[steady grail]

technique can help till a certain point, the rest is all about feeling the timing in your soul

[supple raptor]

I'll just leave that to the pros

[steady grail]

and no logic or theory can explain it

[gusty palm]

practice, music theory

starting early definitely helps since its like learning a language

[steady grail]

trust me, when i touched the piano, not even after a week, i could do things without theory and i'm very shit in technique

it just comes from the inside

[gusty palm]

sorry but i dont believe u

[steady grail]

technique just helps rationalize some feelings

but the rest is instinct

[gusty palm]

chords sound good and pleasing because they sit at mathematically simple multiples that are easy for our brain to recognize

[steady grail]

looks like that bt file didn't load some custom nodes

[steady grail]

yeah true, but the timing is something that comes out naturally

[supple raptor]

did you use your old A1111 venv or the one in your Comfy directory?

[steady grail]

i thinnk have something to do with the Heart beat, but i don't know

the old one

[gusty palm]

sorry but reaction timing and rhythm are learned

[steady grail]

for me, it was natural

0.0 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\FreeU_Advanced
0.0 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\PPF_Noise_ComfyUI
0.0 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_GradientDeepShrink
0.0 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_ADV_CLIP_emb
0.0 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_TiledKSampler
0.1 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_UltimateSDUpscale
0.1 seconds (IMPORT FAILED): D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_essentials
0.1 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_SeeCoder
0.1 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_IPAdapter_plus
0.1 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\Derfuu_ComfyUI_ModdedNodes
0.1 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyMath
0.1 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-OpenPose-Editor
0.2 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-Custom-Scripts
0.2 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\rgthree-comfy
0.3 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-Inspire-Pack
0.5 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-WD14-Tagger
0.5 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-Workflow-Component
1.9 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui_controlnet_aux
2.4 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-Manager
7.4 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_tinyterraNodes
11.5 seconds: D:\Git\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-Impact-Pack

[supple raptor]

ok, lets's get xfomers installed in your comfy venv insted, I have a feeling the A1111 one is just too out of date.

[steady grail]

yeah

it may be so

was just an experiment

[nova remnant]

What kinda of stuff do you play @steady grail on the piano

I have a cheap one I tried for a time

I have a decent ukulele, it's more what I can play I'm noob

[steady grail]

for now i learned some classical and an anime music on the classic side

the first song i learned was very easy, perfect of Ed Sheeran

that i learned thanatos of evangelion

[supple raptor]

from the comfy directory, click the adress bar and type cmd then hit enter to open a cmd window.
in the cmd window type venv\scripts\activate and hit enter to activate your venv.
next type pip list and copy/paste the results between triple backticks like ```list``` so I can see which torch version you're using.

[steady grail]

now i'm learning arabesque and glassy sky of tokio goul

[nova remnant]

classical

You can throw away anime

[supple raptor]

meanwhile, i'm just learning japanese

[supple raptor]

nyo. anime has some excellent classical music in it.

[steady grail]

try to listen to thanatos and than come back

[nova remnant]

You might be right but still

https://tenor.com/view/ace-attorney-phoenix-wright-objection-capcom-gif-20965907

:P

[steady grail]

venvscriptsactivate: command not found

[supple raptor]

did you include the slashes?

[steady grail]

yep

i'm using the portable versiion

maybe that can influence the result?

[supple raptor]

I wonder if it uses conda.

I hate conda

[steady grail]

conda is not for linux?

it should run in pyton

[supple raptor]

Do you have a folder inside your comfy directory called venv?

[steady grail]

no

[supple raptor]

conda is on both linux and windows.

fuck.

[steady grail]

ah, didn't know

[supple raptor]

let me look at the comfy installer real quick to see if I can figure out what's going on.

[nova remnant]

nubby diving into the matrix

[steady grail]

eheheh

i do the same thing, but he is more expert

[supple raptor]

I hate all of the portables and oneclick installers I've come across in the ML space. They're all garbage.

[steady grail]

uahaha

i didn't want my precedent istallations to mess with my gens

https://www.youtube.com/watch?v=Abf8w-EGGcw

tell me this is not a piece worth learning

every time i play it, i get gose bumps

and this, compared to arabesque is still easy

@supple raptor i found this:
Xformers is no longer shipped with the standalones because most of the functionality is a part of pytorch now so it's not needed anymore.

[supple raptor]

yeah, it's talking about SDP which is lies told by people with 24gb+ vram. it's a lot better than other options, but xfomers is king for efficiency.

[steady grail]

well, i'll try installing manually

shouldn't be too hard

[supple raptor]

yeah, I recommend manual install. this portable zip is above my paygrade. i'm no dev and can't decipher how to make it function like normal.

[steady grail]

eheheeh

no problem this my bread

i do this for fun

weird, xformers is already inside the directory

xformers-0.0.21.dist-info

that's exactly the distr compatible with my gpu

but comfyui loads Using xformers cross attention

[supple raptor]

This is quite good even without having watched evangelion before. Since you're into piano, I can't help by recommend you check out basically anything performed by Keiko. She's part of the FFXIV music team. This recording has a bit of static, but it's still a great performance imo: https://www.youtube.com/watch?v=SBBPlFP6lgg

[steady grail]

yeah, teach wise doesn't looks too hard

nothing prohibitive like arabesque, that song is killing me

[supple raptor]

perfect, that should significantly reduce your vram usage

[steady grail]

that explains a lot

it's very calm and emotive

[supple raptor]

here's a better recording of Keiko with some outstanding vocals from Amanda Achen on top. https://www.youtube.com/watch?v=-wxsFNHRu_I

[steady grail]

not stressful at all

i like it

i'll add to my repertory

[dull jackal]

Let me guess---a WILD SPAMMER

[supple raptor]

Hours ago -_-

[steady grail]

a scammer?

[native halo]

yes

[steady grail]

i should move to such a relaxing piece, i'm stressing myself to too with such high end pieces

damn, that voice is very imponent

[supple raptor]

Here's one more of my favorites of Keiko with Amanda Achen. https://www.youtube.com/watch?v=YXyDGhjG8nw

[supple raptor]

Amanda Achen is a gd goddess.

[steady grail]

i'm conflicted if i should play with prompts, or with the workflow XD

anyway i made a funny finding, for how are my settings of comfyui, i don't even need --lowvram

now i'm generating even faster

hoping detailer and other settings will not push the delicate balance of the workflow

[supple raptor]

@steady grail do you like sad anime?

[steady grail]

i like anything done well

[supple raptor]

Have you watched Your Lie In April yet?

[steady grail]

my favorite anime is clannad and i fallen in love with kanon the first time

yes, i love it

i learned so much from it. to love the moment

because what will love now, will change tomorrow

[supple raptor]

It's in my top 2 favorites. Can't pick between that and Grimgar for my favorite.

[steady grail]

Grimgar was fantastic, but an incomplete story

it show how to say farewell to what we loved

[supple raptor]

yeah, I''m very sad it hasn't gotten more seasons

[steady grail]

yep

the first isekai really well done

it deserved more love

[supple raptor]

It's part of why I want to learn japanese, so I can read the novels in the original language.

[steady grail]

yeah, i learned english for that same reason

i feel in love with the queen and some games and i wanted to learn what they said

fire embel, final fantasy

at school i was terrible in english, i learned to read it and a little to speak it, reading about what i feel in love

[supple raptor]

What's your first language? You're proficient enough at english now that I wouldn't have guessed that it wan't your first.

[steady grail]

the queen, aerosmith, stratovarius, sonata arctica

final fantasy, fire emblem, tales saga, metal gear

italian

the power of love

now i read the lord of the ring in english XD

[supple raptor]

Oh neat. Italian really is a lovely sounding language

[steady grail]

yeah, the mother of the classical arts

after latin and Greek, i mean

my love for classical art is an heredity that come from the past of my country

[supple raptor]

It's the most pleasant sounding European language in my opinion. English is my only fluent language and know very little of german and spanish

[steady grail]

when i studied dante at school, i was literally mind blown

english is a very essential language

like with japanese, every context have his word

[supple raptor]

I feel blessed for it to have been my first language. It's definitely as close as we have to a universal language. Though it does feel bland to me, probably largely because it is my primary language.

[steady grail]

italian lacks a word for many situations, but according to the feeling of the subject, it have a synonymous for that word

[nova remnant]

you guys listening to all those fancy songs

[steady grail]

so, you can say the same thing, but you feel the gravity of the context, from the synonym used

[nova remnant]

meanwhile im just screaming murderer

https://tenor.com/view/metal-parrot-headbanging-gif-17332910

[steady grail]

uahahah

[supple raptor]

I kind of enjoy that about some languages because it sort of forces more feeling in the spoken tone rather than relying on just the language to convey an emotion.

[native halo]

when u watchin Death Note and listen the op and ed

[steady grail]

eheheh

[gusty palm]

i posted a music in offtopic yesterday, idk if its good or not i know its not great but i cant seem to judge anything i make myself

[nova remnant]

I have no idea what you just said

[gusty palm]

ive been trying to figure out turntablism scratching without a turntable

[nova remnant]

But I think you're right, you're very wise, steven

(wise, bodysuit:2.5)

[supple raptor]

Have you not watched Death Note?

[steady grail]

his loss XD

[nova remnant]

its so old

I was a kiid when I watched things on tv

and then cs 1.6 happened

I didn't do much else lmao

[supple raptor]

its still gud. I watched it like 3 years ago. for the first time.

It's one of the quintessential animes imo.

[nova remnant]

(Animes are not necessary)

[steady grail]

you should watch stuff from 1998 2004

[nova remnant]

When did you guys first have actual internet

[steady grail]

and than go on

the best years in anime stuff

[supple raptor]

proof that opinions can be wrong

[native halo]

around 2000s

[nova remnant]

((Touhou not very good))

https://tenor.com/view/cat-running-away-escape-getaway-bye-gif-16631286

[steady grail]

touhou is not born as an anime, but as a franchise

[native halo]

https://tenor.com/view/pepe-meme-bullets-wtf-gif-7499092

[supple raptor]

https://tenor.com/view/tic-elder-sister-plastic-neesan-slam-smash-wall-gif-5391652

[nova remnant]

@steady grail https://www.youtube.com/watch?v=vKQ4kV5vWww play this next on piano (somehow )

[gusty palm]

i wanna watch envinleglion

[steady grail]

the video is blocked

[native halo]

i always re-watch that from time to time but only the original series not rebuild

[nova remnant]

What I know about touhou: Nubby posts it all day, frills, hat

[nova remnant]

o: really

[supple raptor]

I got internet for the first time in 1997 and had 56k dialup until I moved out of my parent's house in 2009.

[nova remnant]

not enough neck!

[gusty palm]

lol

oo i remember dial up

[steady grail]

touhou was born as a mean to learn programming of a single person

the first touhou were very stupid arcade games

[native halo]

still remember when all anime we could download were 360p avi files that looked like crap

[supple raptor]

tbh I've never played a touhou game, but the characters are great.

[steady grail]

and than i invested in musical composition skill and than in bullet rain games

[gusty palm]

i wish lost word had some bullet hell elements

the art is good though

and music

[steady grail]

the anime weren't done by such person, but by fans of this person

[gusty palm]

i remember limewire or napster downloading mislabeled mp3s

[native halo]

yea thats where almost everyone got anime from

[gusty palm]

i wish i could go back, and release my own music labeled as like 'metallica techno remix whatever'

[supple raptor]

through the entire napster and limewire lifespan I had to spend 4+ hours to downlod a single mp3.

[steady grail]

eheheh

[gusty palm]

some of the electronic artists i still follow i found origianally on those days

[supple raptor]

I have a 1TB archive from those days.

[steady grail]

i remember i crushed so many 56k

after losing connection

[native halo]

and then u realized it was a soulja boy song he just changed names of popular songs to get more downloads for his mixtape

[gusty palm]

yes id queue up like a whole bunch of downloads, and just leave it running overv night

and wake up to a couple new songs being downloaded lol

[gusty palm]

exactly

[steady grail]

eheheh

[gusty palm]

i dont remember if it read id3 tags or not or if it just went by filename

that was the trick tho

[supple raptor]

id3 tags were visible in limewire if you looked at the details before downloading. I don't believe they were in early napster, but that might have changed towards the end.

[steady grail]

this remind me of megaman for gba XD

that saga was glorious

[gusty palm]

ive been trying to come up with a p2p social network

or some other decentralized way to make it work

[native halo]

i remember some p2p programs had chats

[steady grail]

chats and social are different

we have 4chan on thor

not so social, but decentralized enough

[gusty palm]

i remember ello.co coming out i think it was advertised as decentralized

id like something where each user has like, a local apache server, and then a client that aggregates data from each server in a feed

idk how it would work without a central server though yet so idk

but like an rss feed

meets myspace profiles

something went wrong

[gusty palm]

thats pretty id love a big poster of that

i wish i had like a massive printer my walls would be covered

[steady grail]

lol

[dull jackal]

Yeah, well, slerp

[dull jackal]

Beautiful

[dull jackal]

Can always use Walmart, etc

[steady grail]

soushiki is a beautiful model

can do very beautiful bg

[wispy canopy]

my model folder might be a bit big...

have to copy it off a faulty disk

[gusty palm]

feels

guess whive drive my auto111 is in

i need to find a way to free up space

[whole portal]

https://civitai.com/models/113227/makotoshinkaistyle

i need some better model than this one

this one works with 1 and a half samplers

and gives ____ with others

i need 1.5 Makoto Shinkai style model that works good

[gusty palm]

ive been wanting to do that too

[dull jackal]

I did it recently, when I did the https://www.youtube.com/watch?v=tVte7f14Pkc video

and I gotta say that the cards turned out pretty well

[whole portal]

not exactly what i wanted, but it's nice

[inner summit]

any recommendations for a 1.5 or xl model that would be good to train a character like this on

the ones I've tried have a lot of trouble with the face

[whole portal]

much better

[keen hollow]

yo

what's cookin'?

[gentle musk]

SDXL Turbo Lora, 3 steps, 4 seconds

that's with an ip adapter as well

[gentle musk]

Turbo is pretty good for img2img, this is 3 steps

No prompt so it just copied it

[nova remnant]

using the base model or fine tuned?

[gentle musk]

CounterfeitXL model with the SDXL Turbo LoRA

[nova remnant]

oh

you merged them

[gentle musk]

Someone made an SDXL Turbo LoRA

And tbh it seems better than the base model, if you just want to make images quickly

[nova remnant]

share the link 🙂

[gentle musk]

https://civitai.com/models/215485

[nova remnant]

thanks

did you look at this one https://civitai.com/models/215418?modelVersionId=246043

[gentle musk]

Yeah that's a merge. The Lora is cool because you can use it on any model and get 3 - 4 step images

[nova remnant]

you mean LCM LoRA?

[gentle musk]

No Turbo LoRA, I just linked it.

[nova remnant]

im generating some basic images on sdxl .. with trained models but not using lcm lora or turbo at present

the context is very nice

a beautiful girl sitting by a large window in her living room looking outward, contemplating,

[nova remnant]

ok

this is very interesting

i didnt know turbo lora was a thing

[gentle musk]

This prompt with 3 steps on the Turbo Lora

[nova remnant]

nice ....

you are using sdxl right?

[gentle musk]

yeah

[nova remnant]

ok im downlading that lora.. curious to see how it goes with other sdxl model collections

[gentle musk]

Seems to work pretty well. You do need to use Kohya Deep Shrink for larger images, as it does the same duplication that Turbo would do over 512x512

[wispy canopy]

people figured how to change the turbo model into a lora?

[gentle musk]

Apparently so

[nova remnant]

im using 768x1024 for now

[whole portal]

#🏞｜general-with-images message

[nova remnant]

768

i'll have to use deepshrink if i went over 768

[nova remnant]

any key difference between turbo lora and lcm lora that you'd prefer one over the other?

[whole portal]

turbo should be quicker but with lower quality

lcm give same quality as "normal" sampler with less steps

[nova remnant]

ok

i agree too

[whole portal]

check this #🏞｜general-with-images message

[nova remnant]

i just tested some images with turbo lora they are pretty close in their outputs

[whole portal]

lcm lora + sdxl renderer lora + lcm sampler

i am using 1.5 model as it is sdxl 🙂

no deformed hands, long necks... with 768x1024

[nova remnant]

yeah

[whole portal]

and quality is insane

with cetusMix this time

same prompt, same settings

Time taken: 1 min. 24.0 sec. on M1 mac

[nova remnant]

you know there are some pretty decent devs making newer version of sd1.5 models

i feel like i should go back to sd1.5

[whole portal]

try this lora https://civitai.com/models/171159

but i like the picture on liblib more 🙂

[whole portal]

i am not sure if this image has metadata or not, so lora:LCM_LoRA_Weights_SD15:0.5 lora:SDXLrender_v2.0:0.5

Steps: 8
Sampler: LCM
CFG scale: 1.7
Size: 768x1024

cfg or 2 usugally gives overburned images with lcm, especially with anime models

[nova remnant]

i think you might be able to increase the lcm lora weigh to 0.7

[whole portal]

then you need to lover lcm

which ignores some things in prompt

[nova remnant]

steps?

[gentle musk]

I could never get the LCM Lora to work on steps as low as 3, it always needed 6+

[whole portal]

i recommend 8 steps

[nova remnant]

depends on the sampling method

you'll get very overcooked image with weight 0.7 if used with other sampler than lcm

when i used it with dpm samplers i had to lower the lcm lora weight to 0.3

[whole portal]

this math usually works for me strength = 1.125 / cfg

so, if cfg is 1.5, strengt should be 0.75

cfg 1.6 str 0.7

cfg 1.9 str 0.6

but this is for realistic models

for anime should be lower

let me do the math

probably something like strength = 0.9 (or even 0.85) / cfg

my mac is slow for SD, so i need a lot of time for testing 😦