🍥｜anime

[bitter latch]

calvin got changed.. again

[brittle rampart]

that a not Calvin😀

[bitter latch]

nope

[stable crane]

this is really cool

[brittle rampart]

very atmospheric

[bitter latch]

hmm a furry after a convention?

[bitter latch]

no prompts

ok going to bed. thud

[brittle rampart]

g`night

[gusty palm]

jumbo/aesthetics/general_aesthetics has some interesting tokens lol

Salvagepunk
Sandalpunk
Sanriocore
Scene
Schizowave

[brittle rampart]

last - the best🤣

[brittle rampart]

[Tiled VAE]: the input size is tiny and unnecessary to tile. what this is mean and where i can correct it?

[gusty palm]

i think it means the rendered image is smaller than the tile size

so if tile size is 912 and the image is 512x theres no reason ti split it into tiles

but if the image was like 2048 it would make 4 tiles

im not 100% sure but thats usually how tiled rendereing works in blender

[brittle rampart]

which side(W or H) ?

exp: my im 768x512 and i am try upscale it 2x

[gusty palm]

i think it tries to make tiles fof 1;1 aspect ratio, but it adjusts

[gusty palm]

any thing above the tile size would create a new tile, anything under is 1 tile

[supple raptor]

Yep you got it right

[gusty palm]

just like how it works in blender then uwu

[brittle rampart]

then what parameters should I set in Tiled VAE Encod/Decod for best rezult?

now i set 768 and 64

[gusty palm]

these work for me

tile batch size is the main one ,

not for results but batch size makes it render more tiles at once

overlap im not too sure about and these are probably not optimal

[white zenith]

colorful mess

[brittle rampart]

fumu... i must set true diffusion tiled also? not only Tiled VAE?

[supple raptor]

You don't have to, no. They can each be used separately or together. For Tiled VAE, you want to find the largest tilesize you can use at the batch size you run, without running out of VRAM. That way it will only tile when necessary and will still prevent OoM errors or severe slowdowns due to swapping to system RAM.

[gusty palm]

an i think tiled diffusion only comes into effect if youre rendering images larger than the tile size, by splitting the image to save memory

[brittle rampart]

but sharing can increase generation rate?

[gusty palm]

actually im confused im not sure what tiled diffusion tile size represents

[brittle rampart]

my constant problem of lack of Vram it really solves very well. previously I couldn't use upscaler at all

[gusty palm]

blender only uses one layer of tiling to render

i dont wanna spread misinformation

[supple raptor]

Stable diffusion works with what are called "latents" which is essentially an 8x8 block of pixels. Tiled diffusions measures in latents so you would multiply it's tile size by 8 to get the size in pixels. VAE is what translates between latents and pixels.

[gusty palm]

ah ok

[brittle rampart]

all values must be simply multiples of 8

[white zenith]

-16 minutes of gen time to wrong vae loaded 😭

[supple raptor]

I think that's exactly why the dev decided to use latent values for tiled diffusion. Since it's already in latents, the values for that one don't need to be multiples of 8.

[brittle rampart]

forgive my lack of intelligence, but still, in the end, I still didn’t understand how will using Tiles help me increase the speed or quality of the resulting images? besides eliminating errors related to the lack of Vram?

[brittle rampart]

of course a good method in Tiles is to get images based on regions!

[supple raptor]

It only increases speed by preventing the gen from exceeding your VRAM since that will cause things to be swapped into system RAM which is painfully slow. Aside from that, if you can gen without tiling without running out of VRAM, it will be faster to do so.

[brittle rampart]

now I understand. In my case, using Upscaler without tiling will not work.

[supple raptor]

Yeah, there are limits to what can be swapped to system ram, so depending on which part of the process it's at, it will either throw an error or swap something to system RAM

The worst part is that if you're not using Tiled VAE then the most likely part to get an OoM error is on the VAE step which is at the end of the generation.

[gusty palm]

it feels like using tiled diffusion and vae u could make images a s large as u want

[brittle rampart]

in my case, the error occurred either immediately or in the middle during when the Upscaling process os run

[gusty palm]

just take longer cuz its done in batches

i know its for sure doing something in my img2imgs

[brittle rampart]

exactly!

[white zenith]

vae requesting 200gb of ram to decode... pain

[gusty palm]

im already conditioned to wait a few minutes per frame in blender

200gb of ram

😮

[supple raptor]

There are some limits, that I don't fully understand the reason for, but within reason, yeah. I've made some that were over 16k resolution.

[brittle rampart]

I could never get images right away even in FullHD resolution

[white zenith]

it just ooms if you increase the size too much, on DirectML comfyui it does just eat ram like this (amd pain)

[gusty palm]

ive seen your wide renders i love em

[white zenith]

I think for a base image it peaks at close to 60gb for normal sized stuff

[gusty palm]

got me all inspired

[brittle rampart]

my!?

[gusty palm]

the sygil ui i was using before had pretty much no img2img capability

nubby

but i like youres too uwu i havent seen your wide ones thoo

ive always been a fan of desktop wallappers

[supple raptor]

Are you familiar with the Krita extension? It's excellent for inpainting which can be very helpful on big gens.

[gusty palm]

i saw it in the list but wasnt sure if id be able to figure it out

[white zenith]

miku

[gusty palm]

i saw an extension called asymetric tiling or something that sounded like it might be useful for seamlessly merging 2 images

[brittle rampart]

but maybe she bit smiling?

[gusty palm]

i havent found anything to test it on though

[white zenith]

I think the upscaler does a kinda ok job

[brittle rampart]

very interesting style !

[gloomy lodge]

something other than cute girls is from from time to time

[brittle rampart]

Well, I'm more likely not about the content, but about the design

[gusty palm]

https://civitai.com/models/117359/mbbxl-alpha is there any special requirement to use these or just more vram

[steady grail]

how do you see if the gen is using ram instead of vram? i don't get out or memory any more, but i would like to prevent slow downs.

[steady grail]

i love sunsents

[hearty raven]

Get Linux, and just allocate swap memory

[steady grail]

i could use the emulator, but it's boring

[hearty raven]

yeah docker is a good solution too

[supple raptor]

If you're interested in it, I can take some screenshots of how to use it. With how quick you've picked up on everything else, I don't think it'll be a problem for you.

[hearty raven]

real @supple raptor ?

[supple raptor]

You'll know because your generation speed will be like 10x slower.

[steady grail]

every time i use high res fix and x2 upscaling, my gen became 10 times slower

so it's probable it's leaking on ram

[supple raptor]

I think I'm real

[hearty raven]

isnt that the case for everyone?

[brittle rampart]

6.02s/it - it is slow?

[supple raptor]

Well, increasing resolution will also reduce gen speed, but usually it'll be more like half speed assuming a 2x upscale.

[steady grail]

so, i'm leaking for sure

[supple raptor]

Probably, I think the latest Nvidia driver has an issue that causes offloading to system ram when it shouldn't.

[steady grail]

i get something like 15/30 sec without fix, but with fix, 10/15 min

but even with tiled vae, it's pretty slow. the difference is almost unnoticiable

very interesting

[brittle rampart]

15 sec.. nice i am get 1+min

[supple raptor]

Yeah, that definitely sounds like it's swapping.

[brittle rampart]

with fix i am get 10-12 min

[steady grail]

i'm using an old branch of stable diffusion, where xformers was better performing for my gcard

i can even push without half\no vae, without error ❤️

[white zenith]

cfg too low

[steady grail]

i love this style

[white zenith]

you can clearly see the blocks where it does an upscale

although they look really odd once they get denoised

[brittle rampart]

btw. guys... tell me if i am use Tileing VAE i can remove from my bat-file any of these parameters as not needed?

--lowvram --always-batch-cond-uncond --opt-split-attention --disable-nan-check --xformers --no-half

[steady grail]

i use --always-batch-cond-uncond --opt-split-attention --xformers --medvram

[supple raptor]

Which GPU do you have?

[steady grail]

remove -disable-nan-check

[brittle rampart]

GF GTX 1660 6 Gb

[supple raptor]

--medvram would probably be better for you than --lowvram

[brittle rampart]

disable-nan-check --no-half this is ?

[supple raptor]

Keep --no-half since 16xx series doesn't support fp16.

[white zenith]

settings for XL are kinda cooked but it worked

[brittle rampart]

for me better left

[supple raptor]

--disable-nan-check gives you a black image instead of an error when either the UNET or VAE fail math class. That one shouldn't be necessary, especially since you won't be running at fp16.

[brittle rampart]

ok now i try..

[steady grail]

nubby, i use nvidia gforce gtx 1650 without no-half, using an old xformers, why this is possible?

[brittle rampart]

--disable-nan-check -then i can delete too?

[steady grail]

this is a thing i can't still understand, maybe that old xformers, don't generate half-vae?

[supple raptor]

Not sure, probably doing some up casting shenanigans.

[steady grail]

cool

[steady grail]

xl?

[white zenith]

Generate 768 image with SDXL-> latent upscale -> "refine" with 1.5 model at 1160 to get closer to anime

[steady grail]

oh, you are running sdxl

i would love to try sdxl on a 4gvram for fun 😦

do you have guide to run it?

[white zenith]

no, using directML (amd card) so it hogs ram like an absolute monster

[steady grail]

oh, you re running it on ram and not vram?

[white zenith]

not too sure how well 4gb card would work, I know 1.5 ran even on like 2gb but xl likes large images+is way bigger

[steady grail]

that's what i want to experiment, i like to do wizzardy

[white zenith]

I mean gl I guess lol

I dont have enough ram so its chewing up my ssd as a pagefile

1tb+ written just tinkering today lol

[steady grail]

eheheeh

your gc is a 16 gb?

[white zenith]

6900xt yeah

this one is just weird

[steady grail]

and you leaked 34 more gb on ssd

[white zenith]

its mostly whenever the XL vae loads it starts to eat up ram

[steady grail]

it's a monster O.O

i can perform some trick i guess, but wow

my ssd is almost empty, but... even with some touch, 150 gb of ssd is even enough on a 4g?

would be funny if i could push something like this lol

[white zenith]

idk will be painfully slow if it cant load the entire model to vram, like 10-20X slower depending on how much it has to use system ram

and if you run out of ram 💀 💀

[steady grail]

this time, i can even risk melting something...

lol

[white zenith]

its not gonna melt its just gonna be idle like 99% of the time because loading from system ram instead of vram is so painfuly slow

[steady grail]

last time i messed with python, i got some weird results... lol

[steady grail]

i had to let the pc cool down for some minutes, that was nuts XD

xl is very creative

i love ai, i can't stop having fun with them

[brittle rampart]

style\(medium\)
explain if not difficult. and what other options besides medium
can be used?

[steady grail]

it's a prompt?

[brittle rampart]

yes

[steady grail]

maybe a wild card?

i would do the same

where did you get this prompt?

[brittle rampart]

exmp: (watercolor_style\(medium\):1.4)

[queen void]

Well sdxl works perfect on a 3060 mobile/laptop. The lowest gpu I tried yet.

6gb vram

[steady grail]

a giant

[bitter latch]

yeah not sure why

[steady grail]

@hearty raven

[white zenith]

headwings

[bitter latch]

interesting

[white zenith]

is this 1.5? the deepdream prompt works much better on xl

[bitter latch]

cant' seem to get that effect but getting great results with the new merg 😄

[white zenith]

its a weird workflow

[bitter latch]

throw stuff from style pile in there and it just drifts away

like this

1 more and i'm going back to bed

well that got dark

ok I can't leave on that

[white zenith]

6 prompts (clip_G +/-, clip_L +/-, prompting for NAI model) but I wonder what is doing what

[bitter latch]

there we go

[steady grail]

@hearty raven

[bitter latch]

hha

dang it this just one more... sigh

i'm doomed

there

goes to bed

leaves this here...

goes to bed... thud

[steady grail]

lol

[brittle rampart]

why didn't anyone tell me earlier about the "sd-webui-prompt-all-in-one" add-on!!?!?! it's just magical!

[steady grail]

what's that?

[brittle rampart]

replacing the standard prompt input field with additional functionality

[steady grail]

should try it

[brittle rampart]

since English is not my native language, the translation function here is a godsend!

[steady grail]

ohhh

[brittle rampart]

history of used keywords

[steady grail]

totally forgot how it is not knowing basic eng

[brittle rampart]

I'm not talking about basic knowledge of the language

and a certain vocabulary I have

[steady grail]

usually, the words used in prompting are pretty basic

[brittle rampart]

is it? not always like this

common words, yes. but if anything is original

[steady grail]

the models are usually based on booru indicization and google results. for me that i prefer navigate eng sites, everything is pretty intuitive

[brittle rampart]

Again, I'm talking about convenience.

[steady grail]

if you are looking for new prompt, just look for images on booru and google reserch

translations may not be faithful whith what you hope to obtain

the word superb is a fantastic example. some use it hoping to get a better image, but the model riconize it as the car. why that? because if you write superb on gogle, most result are about the car

cool, right?

[brittle rampart]

therefore, I can replace it with many words from my native language, which will be translated differently and about the spelling of which in English I may not know, agree ..

speaking of "superb" until the case when I encountered this, I did not know what it meant in prompts. but I saw it everywhere.

[steady grail]

probably they copied from someone else, or the model was trained differently

that was an exquisite example of how models works

do you have https://github.com/DominikDoom/a1111-sd-webui-tagcomplete ?

[brittle rampart]

nope

[steady grail]

this help to find booru synonyms. will be pretty useful together with translations

[brittle rampart]

thanks try it now. and you know extention for highlighting prompt?

[steady grail]

with different colors?

i remember there is something like that out there, but never used it. but i can look for it

[brittle rampart]

yes

[brittle rampart]

please

[steady grail]

i'm bad at drawing knees lol

[white zenith]

super cozy

[steady grail]

cute

[brittle rampart]

fingers is @$##.....

[steady grail]

nothing, i can't find it anywhere

[white zenith]

the upscaling artifacts tho

[brittle rampart]

ok.

[steady grail]

it may be obsolete

[brittle rampart]

don`t worry

[steady grail]

no more compatible

[gusty palm]

ill try to find something i can use krita for and take u up on that
Im better with inkscape and glimpse

has anyone seen those youtube channels with the like video fottage from the 18/1900s restored and stuff with ai

i bet video games in the future will use like a camera in an ai environment or something

java minecraft used to have some command line args that let u set exacly how much ram the game was allowed to allocate that would be useful

-medvram seems to only use like 4gb itd be cool if ther was a way to tell it to use 6 and no more and to just put everything past that into a tile queue

[gusty palm]

i actually havent been using medvram or anything beside xformers

[bitter latch]

not sure where the horns came from..

morning

[gusty palm]

suprise horns

morning

[gusty palm]

evily turns cfg to 16

[bitter latch]

turns his to 18

[gusty palm]

are u goin for like a harley quin

[bitter latch]

ooh burnout

[gusty palm]

i wanna make a sexy bugs bunny

[bitter latch]

thats lola

😄

[gusty palm]

i love these skullgirls looking waifus

truee

[bitter latch]

better

[gusty palm]

oo we both got sorta normal looking hands

[bitter latch]

had to turn it down

one last one then work

[gusty palm]

i bout to go too :<

[bitter latch]

hmm it will not do bugs bunny wtf

[gusty palm]

have a good day

awh

[bitter latch]

later

[elfin bolt]

anyone know how to stop the comfyui nodes to save some ram (like mid stop kill-switch) and then continue the workflow after the ram saved? (I managed to create the image from SDXL base and refiner, but crashes when I tried to put facedetailer nodes since it used too much ram)

[wet elm]

damn i cant use sdxl with my gpu

[vocal elbow]

any prompt for generating specific mask like Demon Slayer: Kimetsu No Yaiba

[vocal elbow]

ah kitsune mask should be the answer

[gloomy scroll]

What gpu you have?

[wet elm]

GTX 1060 6gb

[gloomy scroll]

Have you tried with Comfyui?

@wet elm ^

[nova remnant]

Okey thanks i will try

Is made by ai

[gloomy scroll]

img2img

Probably some cnet

[elfin bolt]

MikuMikudance animation, controlnet-temporalnet, controlnet-softedgehed, stylized lora, low denoising (0.25 - 3) img2img batch, davinci deflicker, flowframes

[wet elm]

no, it seems way too complicated

[wet elm]

but i want to try it

[gloomy scroll]

I have heard people using 6GB GPUs on Comfyui

[gusty palm]

today i will use my gt720

ah nvm i was gonna take a picture of it but its in a box in the other room

[nova remnant]

What is temporalnet and softedgehed

[elfin bolt]

https://huggingface.co/CiaraRowles/TemporalNet2

https://github.com/Mikubill/sd-webui-controlnet

tool to help maintain frame consistency

[summer shuttle]

slime

[nova remnant]

But the characters are created on stable diffusion or on mikumikudance?

[elfin bolt]

MikuMikuDance, Stable diffusion have no capability to create that fluid animation as of now

[nova remnant]

ok thanks

[primal wharf]

I see some people posting big thighs, so here's some skinny and long legs:

[primal wharf]

skinny

[rough oxide]

comfy shared a neat way of making anime images with sdxl; (workflow is on pic)

[rough oxide]

what sdxl makes vs how anime model finishes it off

[keen hollow]

5760x2400

[rough oxide]

sdxl is now a beast at anime

xD

[keen hollow]

using comfy?

[rough oxide]

Yup

u can paste this image into ur comfy to see the workflow, credits go to Comfy himself for the basic setup

i've been tinkering with it a lot

[keen hollow]

I need some comfy training or something. Found it sort of cumbersome. Not sure why I'd choose that over Auto.

[rough oxide]

testing all sorts of things still lol

[keen hollow]

like having to put in nodes to move in and out of latent and stuff.. just really weird workflows

[rough oxide]

this is an upscale workflow

if you notice, end image is 1536x1536

that's why the latent upscaler is there

[keen hollow]

I've been using Tiled Diffusion/VAE a lot recently, too, which I don't even know if you can do in comfy...

latent upscaling blows though. 😄

[rough oxide]

but

it's because

it's finished within a second model

for just a regular comfy setup

for stable diffusion XL, no tricks or anything, here's my workflow

*you need the new sdxl/refiner models with baked in vae

that just released

[keen hollow]

haven't messed with SDXL yet either.. the whole add-on steps for the refiner just seem like too much hassle

[rough oxide]

nuuuu

that part is simple lol

[keen hollow]

gotta set independent samples, 80/20 rule, and all kinds of nonsense

[rough oxide]

yes

lol

but i did all that

[keen hollow]

I do like that, once you have a workflow set up in comfy, you can create templates

[rough oxide]

if you use my workflow it's already hyper-optimized

all fits in the smol starting bounding box square

xD

anything that isn't necessary is outside of it

[keen hollow]

can you use tiled diffusion?

[rough oxide]

never tried it

afaik

aside from adetailer

[keen hollow]

love me some adetailer

[rough oxide]

anything possible within a1111 is possible within comfy, usually more efficient

but the real q is do you 'really' need it

i'll post my latest 2 outputs in a sec

[keen hollow]

This: https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111

[rough oxide]

look at face on one on left, is adetailer REAAALLY necessary? xD

...totally messed up in other palces tho

lol

[keen hollow]

I like adetailer if I'm running a lower res set

[rough oxide]

fair

[keen hollow]

I am often trying to get to 3840x1600 (or even a little larger at same ratio)

but I'll run sets to get a flavor first

[rough oxide]

i guess for me adetailer was nice if

there was someone really far back in the background

then it actually made the face presentable

but uhhh

i'll never go back to a1111 most likely

down to just

optimization/efficiency

comfy UI is literally much faster for me

• cuda-malloc

a1111 doesn't have a setting for that afaik

[keen hollow]

Ran denoise too high but it ended up being kinda cool. The details are nuts. 5760x2400.

[brittle rampart]

I wonder if you get paid to advertise or do you advertise SDXL for free? if not, you should raise the issue.

[keen hollow]

comfy definitely loaded a lot faster for me .. didn't notice much in regard to generation

[rough oxide]

...wat?

[rough oxide]

the cuda-malloc fix added to comfy makes it much more efficient, less vram usage, faster gens. NVIDIA's recent drivers introduced a bug and a1111 hasn't fixed it through their UI afaik

it's enough of a fix that i can run 3 batch size when i can only do 1-2 normally

like NVIDIA drivers REALLY messed somethin up

[brittle rampart]

forget

[keen hollow]

ooof... I'm definitely on older drivers

[rough oxide]

yea, it's good if ur on older drivers

any driver prior to 531 i think

is fine

[keen hollow]

i dont even wanna know what I'm on

[supple raptor]

The bug has already been acknowledged by Nvidia and they said a fix is in the works.

[rough oxide]

Oooh! that's good news.

[keen hollow]

Driver Version: 516.94

oof

[keen hollow]

are people sharing Comfy templates somewhere?

[rough oxide]

i have a template for anime and one for sdxl inference

[elfin bolt]

you can literally just drag and drop png from comfyui to see someone's workflow

[rough oxide]

that too

For the most part

there's a disable metadata option now

but it'd work on any of my pics lol

cuuuute

[keen hollow]

I often use a workflow where I'll generate a prompt, or prompts, through my entire modelset.

Another is taking a 960x400 image, hires to 1920x800, then img2img with tiled diffusion to 3840x1600.

[rough oxide]

i always gen 1024x1024 :<

first

though the second pass gen is a 1536x1536

for these ones

[elfin bolt]

I've seen your workflow, it's t2i SDXL > i2i anime model, is that right?

at denoising 0.5

[rough oxide]

yea; it's modifed from a workflow comfy posted

i think i've got it where i want it to be now

[elfin bolt]

Okay, imo, I'm still hoping that we don't need refiner or postprocessing and just straight good output from SDXL based ckpt

[rough oxide]

(forgot to describe an expression on the i2i side here XD)

no the refiner never comes into play in this one

infact i've noticed refiner usually makes anime outputs worse so i didn't even include it

[elfin bolt]

Ah yeah, but we still need anime based model

[jagged wasp]

doesn't look much different from good 1.5 outputs

[rough oxide]

The thing is, with SDXL making the base image, it's a lot better

SDXL has more detailed backgrounds, better anatomy, understands more concepts

Also SDXL has better expressions, maybe not facial structure and way of showing it always

but you can get a much wider range of expressions with SDXL

and then transfer those to the anime model as well to work with

still working on this prompt but another output

i have, occasionally, been able to get really good exaggerated expressions in a anime model mix

however

[rough oxide]

most of the times exaggerated expressions just don't come out well afaik

but sdxl can start you on a REALLY exaggerated expression lol

that's really good/high quality

[jagged wasp]

I guess the biggest improvements will come when ppl start releasing fine tunes, I'm still not "wow" over sdxl base

[keen hollow]

I like this one. Neat colors.

[elfin bolt]

this would be very long.. great anime model didnt even exists until NAI leak happened

Waifu Diffusion were just okay at that time

and Anythingv3 were based on tremendous NAI finetuning

then the derivatives started to come based on NAI leaks

[swift lava]

It may not take as long because the base model seems far better than 1.5 was base.

Less work would have to be done to train it.

[jagged wasp]

the way things have been I'm betting the time to fine tunes will be a lot shorter this time around

[elfin bolt]

Yeah, but so far the animediffusion from lykon and other loras is still very stiff

[rough oxide]

sdxl still starts the base waaaay better than any of the previous models can

i can't even get red auras to appear most of the time in any anime model

[elfin bolt]

probably 1 months before the community can catch up with the current mixes of sd 1.5

[rough oxide]

because sdxl understands soooo many more concepts

it's soooo much better as a base model imo

sdxl is without fail, always giving me the red aura i ask for

anime models just... don't do that

but if there's one there present already

it'll enhance it

[jagged wasp]

I feel like auras was more of a Lora thing with 1.5

[rough oxide]

welllll i don't use a lot of loras or embeds unless necessary xD

guess that's more of a me problem

but with xl you almost odn't need to

since it has almost any style you could ever want embedded in it

and way easier to call out than 1.5

[elfin bolt]

I'm just waiting for controlnet to come out for SDXL, because that's vital part of my workflow

[jagged wasp]

that's how i'm feeling about XL it's more of a reset to a new base not so much a hugh upgrade out the gate

ppl have pushed 1.5 so far from what it was on release

[rough oxide]

ye

it's AMAZING what 1.5 finetunes were able to do

but i felt like it hit a wall

[pseudo condor]

Neeko 😄

[rough oxide]

just using XL as a base and then applying those old finetunes again as an i2i feels like a massive improvement to me at least

[jagged wasp]

i2i i think is what will get the most return on any base

[rough oxide]

i do not think i'd be able to pull ones in a style like this off without the help of xl as base xD

[brittle rampart]

alas, but to repeat, even close to the softness and style of watercolor, you did not succeed.

[rough oxide]

i wasn't trying to mimic it lol

tbh

this was more of a test if i could get sweating profusely to work again xD

seems like it's working fine overall

just, the idea of purple haired elf i wanted to seeeee

[brittle rampart]

test of what?

[rough oxide]

how well this workflow is working for various themes

but i did use some of the idea of that image which is why i replied lol

[white zenith]

interesting, another person using XL to anime model

[rough oxide]

yea; this is so far the best method i've found. the quality is insane

[white zenith]

the possibility of using 6 prompts on an output is pretty interesting

XL is way better as a base model for gens though, no 2nd hand growing out of a leg or any random garbage like that that you usualy get with older models

these gens are kinda scuffed though since upscaling latent before img2img step introduces a ton of noise, I guess it gens faster though

[rough oxide]

upscaling latent before gen is what i'm doing though and it comes out fine ._.

but i also increased the number of steps the anime model does from where it was originally

so that might be helping it too

[white zenith]

I think this workflow was XL latent-> latent upscale->vae decode->vae endcode-> 1.5 anime model

might work better as like XL->vae decode->vae encode->latent upscale

[rough oxide]

yea, that's what this is doing

vae encode into the upscaler

[white zenith]

https://tenor.com/view/diesofcringe-gif-24606059

[rough oxide]

on mine? change batch_size to 1

i always gen in batches

to save like 3 seconds lol

[bitter latch]

😄

[brittle rampart]

both deer will rust

[bitter latch]

oh dear

oh dear, oh deer, oh dear

and converted

[cold talon]

which model is that I ponder

pretty nice texture

[brittle rampart]

where is the second deer?

[bitter latch]

shrugs

[brittle rampart]

no no no no .. i wont eating her!

[brittle rampart]

OMG!!! 🤣

[supple raptor]

It's so cursed XD

[brittle rampart]

it`s a horror!

[gusty palm]

she borger

[untold glacier]

[gusty palm]

which u think looks best

idk why .4 looks best

to me

[jagged wasp]

hmm pushing a XL base through a 1.5 model like mimi was saying earlier...

[rough oxide]

ye love what it can do, still testing so many ideas lol

[gusty palm]

ok fine ill try xl

[rough oxide]

one day models will be able to do held objects xD

[jagged wasp]

we'll be 3d printing the waifu's by then, for science purposes ofc

[untold glacier]

I'm thinking there must be an issue with my setup somewhere.

ɪ ʜᴀᴛᴇ ᴄᴏᴍғʏᴜɪ

[white zenith]

tbh I dont like the high denoise gens, inherits too much of the "1.5" look

its just the way the lighting is inherited through all the models that ends up having a "look"

[jagged wasp]

i would agree

it's just where we end up once the refines come out you have to sharpen a dull edge to a point and after a while all the sharp edges look similar

[jagged wasp]

it is making some interested images though

[white zenith]

can probably be fixed with MBW merging to find the right model if you tinker enough 💀 (maybe merge an offset noise lora+ a few layers of a heavily stylized model?)

[gloomy scroll]

https://media.discordapp.net/attachments/1019446913268973689/1108357749005369415/c7572bd4-9d9d-44d2-9752-459dd09d03ab.gif

[untold glacier]

Most of my gens end up with big issues, and there's a persistent error that I can't find any info about.

[jagged wasp]

I'm slowing lowering the denoise with each gen

[white zenith]

bruh the arms on that skeleton looking downright delightful

[supple raptor]

https://cdn.discordapp.com/emojis/1134607661640458280.webp?size=64&name=flandrenom&quality=lossless

[rough oxide]

i know what causes this

it's a sampler mismatch on your 2nd step

e.g. say you use dpm++ 2m normal scheduler on first part, but then karras on 2nd, or just outright a diff sampler

use the same scheduler/sampler for all parts

[rough oxide]

you can mess with cfg/noise/etc but sampler/scheduler must be the same

[untold glacier]

It was the same.

[untold glacier]

I... may have found the issue though. Your suggestion led me to a different problem that I found: the amount of total steps was different than the scheduled amount.

[rough oxide]

o i was running through it now to see

[untold glacier]

Making things with comfyui is like all of my tools are now controlled with wet noodles.

It's messy, and there's a lot to keep track of. Everything is connected in strange and unexpected ways.

[steady grail]

comfyui is more efficent than stable diffusion?

[rough oxide]

?

comfyui is a UI to run stable diffusion

[steady grail]

oh, like auto1111

[white zenith]

do a dark summoning ritual

[untold glacier]

I'll take "Reasons why I never want to play Factorio" for 300, Alex.

[steady grail]

so, it's a gui

[gusty palm]

nodes provide a good visual layout of info processiong

[rough oxide]

it might also be the SDE sampler(?)

i switched sampler and it worked but completely diff

[untold glacier]

Currently testing to see if it's fixed now

[steady grail]

works better than ... auto, right?

[untold glacier]

[white zenith]

the skills for reading the sphaghet pile transfer surprisingly well

[rough oxide]

altho might not be that cuz i'm using sde_gpu for the anime stuff

hmmm

[pseudo condor]

more neeko = D

[white zenith]

maybe a tad more denoise

[jagged wasp]

started at .45 denoise and down to .25

[steady grail]

we should connect a match of fatorio to confetti... confyui

[untold glacier]

Lol. Sadly, it didn't seem to help too much. I dunno. 1 = 2M SDE, 2 = EulerA, 3 = Euler

[rough oxide]

Hmmm

It might be

refiner itself

doesn't like SDE

because i had that same issue when i tried using SDE

I think refiner model can't use SDE so you probably don't wanna use SDE for a pure sdxl build

[white zenith]

isnt SDE 2 steps per it or something like that?

[untold glacier]

Euler A/ Euler still have problems though. I dunno.

[rough oxide]

euler/euler a seem fine

just bad seed

[untold glacier]

Hard for me to judge, with my lack of SDXL experience.

[rough oxide]

oh

also

there's a lot u have to change if u wanna change width/height

if not doing 1024x1024

u gotta change all the target/actual width/height nodes on everything too

probably easy to set that up through a primitive if it works

[white zenith]

wider

[gusty palm]

just ran my own selfie thorugh img2img

do not recommend

[gloomy scroll]

With high enough denoise you will turn into waifu

[supple raptor]

everything's a catgirl if you're brave enough your denoise is high enough

[gusty palm]

yeah ><

its a great way to see where the sweet spot for denois is

[gloomy scroll]

1.0 is good place to start when i2i selfie

[slate night]

Happy H✨o✨La✨! to you too!

[gusty palm]

i didnt even think to try that. i tried on liky 0.3 and then .45

[untold glacier]

Uh... that's... not exactly what I asked for.

[untold glacier]

Think I'm finally starting to get a feel for some things with SDXL.

[terse ore]

I'm trying to use the FaceDetailer node in ComfyUI, but it gives the error message that it just skips. It didn't happen before.

[rough oxide]

kakigori? looks nice.

speaking of kakigori what changed within v2?

[supple raptor]

Yep, v2 used v1 as the base with 2 more models merged in using MBW (Awoo7 and Vanilla) along with 2 loras to help bring back some of the toonlike style that was lost by adding those models (Lilliput and ssambatea)

[rough oxide]

oooo, ic. i'll have to try it out

[supple raptor]

The merge with Awoo helped to make the model more compatible with other models (needed to facilitate the merge with vanilla) as well as improving the hands and backgrounds, and Vanilla brought in even more improvements to backgrounds (especially simple ones)

[rough oxide]

loving the way it handles backgrounds in the second one, and toon style is great for the cheerleader thing i'm trying to pull off xD

forgor to switch to simpler neg prompt, now it's better xD

[nova remnant]

reminds me of Archie comics

[bitter latch]

hmmmm

does good on the woman tho doesn't follow prompt on clothing. odd

or background

[hearty raven]

rng'd this cirno

[gusty palm]

poros!

[bitter latch]

just doesn't wanna do that one for some reason odd

[nova remnant]

mine was kind of 3D anime theme

[nova remnant]

so many wasted images cause of messed up fingers

SD fundamentally hasn't yet been able to get the fingers right

[rough oxide]

yeaaaa lol

you can normally inpaint to fix those but i'm too lazy to inpaint tonight

[nova remnant]

im using the adetailer but that's still a hit and miss

but at least adetailer is great for fixing faces

the finger issue is a basic problem with SD engine, can't do much about the gen, after gen edits are different matter

[oak spoke]

[rough oxide]

one thing i learned is never describe fingernails xD

like nail polish & all that

lately having SDXL do the base also makes hands way more accurate too

[nova remnant]

im trying out (badly drawn fingers:1.2)

or (malformed fingers:1.2)

(malformed, distorted fingers:1.2)

(malformed, distorted, interlocked fingers:1.2)

[lean forge]

what model is this

[supple raptor]

Looks like some flavor of AOM, if not, it's some merge consisting mostly of AOM.

[lean forge]

oh wow thanks

[supple raptor]

happy to help. hopefully tekmunki will pop in to let you know specifically which version it is, but at least that will get you the general style

[sullen depot]

how the hell u guys can make that hand looking damn cool

is there any lora or lyco or something else?

[median herald]

whats a lora or lyco?

[sullen depot]

IDK too much but it kinda got trained to make more stuff

Can I get your promt pls

[zinc flax]

counterfeitXL looks nice

[gloomy scroll]

Lora/Locon/Lycoris/Loha/Lokr are files that have been trained with some specific Concept, Character or Style and can be added to prompt like this <lora:NiceAnimeStyle:1.0>

Loras are saved to webui\models\lora\

[cyan burrow]

because I love you all so much, I gift you this

[median herald]

moths are so cute

[median herald]

ill have to dm u cus of word restrictions

[keen hollow]

You switching over to comfy? 🙂 Messing with SDXL?

Let me know if you figure out the Tiled Diffusion workflow. After you helped me set that up I use it a lot. 😄

[cyan burrow]

with some post editing, scaling up and photoshop fixing and adjusting. I changed A into B

[final vault]

sdxl?

[final vault]

which model bro

[gloomy scroll]

Kakigori v2 + add_detail lora

[heady shard]

If seen this image on Twitter. Do you know if its Stable Diffusion? And if yes, do you have an idea of which model it could be?

[lyric zealot]

not the best place to ask this, since this is anime channel

[heady shard]

I thought its like a mix of anything and anime

[lyric zealot]

looks a little more realistic to me. You might wanna try bra model

[heady shard]

yo thanks

[regal pivot]

@heady shard is SD gen and .... no is anime ... go to Civitai any realistic can do it

[cyan burrow]

heh, there's so many models with really good anatomy

and Loras to improve it and make it more realistic

[untold glacier]

But I have been trying SDXL out with it. I hate the workflow. And its overcomplicated nature.

The second Auto updates his build to support SDXL I'm deleting Comfy from my system.

[lyric zealot]

it's actually kinda interesting using comfyui for the first time tho, not gonna lie

i believe the only perk for me, is probably the increased performance when using comfyui

[untold glacier]

True. And it helps one (forcibly) understand what exactly is going on in SD.

[lyric zealot]

true true

[untold glacier]

It just takes away 90% of how and what I usually work with, heh. Hires-fix, i2i upscaling, the ability to look at what keywords an image has without replacing the workflow, Adetailer, tiled diffusion, etc. etc.

[lyric zealot]

true true, im almost certain you can use hires-fix with that tho

but im dumb so don't trust me

[untold glacier]

I haven't seen a workflow yet that has that ability with SDXL.

[cyan burrow]

What's that new webui called that's similar to comfyui but more expansive?

I heard it can assign multiple gpu's to different parts of the generation, or maybe it's just if you're running several images per rotation

[summer shuttle]

img2img thing

[regal pivot]

Has anyone tried that new project .."StableSwarmUI" ... it looks like the horde but I think he is trying to implement more things in it ?? ... I couldn't check it because I just sold my other laptop with a gpu, and in this country you wouldn't see improvements if it wasn't in a local network..

[lyric zealot]

isn't that the new ui by stability itself?

[cyan burrow]

yeah it was stableswarm

I haven't tried it

[lyric zealot]

oh wait, might not be from stability

[cyan burrow]

but it has both nodes and single and multi-gpu

no its not from stability

[lyric zealot]

i think you can literally use collab gpu with the ui

so you can essentially use your own gpu + collab gpu for faster(?) generation?

[cyan burrow]

or two gpus of your own

[regal pivot]

They say that they try to improve the speed by redistributing the stream between the available gpus..

[cyan burrow]

like an intel 16gb+an invidia

the intel 16gb are pretty cheap now

[lyric zealot]

hmmmm

i should go unplug my old computer gpu

[keen hollow]

didn't this already happen?

[keen hollow]

I think it would be the same regardless of whether or not you use SDXL. It's just that in Comfy it's like 4 nodes, and in Auto it's a checkbox. 😛

[untold glacier]

Auto's build can't have the model and the refiner loaded at the same time, no. And regarding the workflow similarities, SDXL has new things to consider regarding latent cropping that I don't fully understand yet with the refiner, and we would need to ability to alter/change that to properly generate images at different ratios. That and they've been pushing specific resolutions that work best (that they've trained with SDXL.) So having those within the UI as presets would be nice.
Some of that can be fixed via extensions, but the refiner issues likely need to be done at a deeper level.

[keen hollow]

Don't you just pass the image through the refiner in img2img?

[untold glacier]

Let me try to find what I'm talking about

[rough oxide]

No.

You pass a still noised image to the refiner to finish the last 20% of steps

Base model does 80% of timesteps, refiner does 20% of timesteps

[untold glacier]

Ah, mimizukari is here to explain, heh.

[rough oxide]

No img2img involved.

[keen hollow]

Got it

[rough oxide]

refiner is not for img2img, and though it can sometimes make something look better

that is an unintended use of it and will usually produce much worse quality

lol

[untold glacier]

Mimi, do you have a link to an article or anything that talks about the changing of the ratio from 1x1, and how it's more involved than 1.x? I was looking at one the other day but I can't seem to find it now. Or maybe I'm remembering things incorrectly.

[rough oxide]

i don't

[keen hollow]

Is this the official workflow people are using in comfy? https://github.com/comfyanonymous/ComfyUI_examples/tree/master/sdxl

[rough oxide]

No; i think most people use Sytan's

[untold glacier]

Alright. Because you have all of these in the workflow you gave me, and I was wondering where you got your info from for them.

[rough oxide]

though Sytan's isn't a full denoise

[keen hollow]

Why his over this?

[rough oxide]

turned those to widgets to save space; i think you generally want target width/height and actual width/height to be the same from what people discussed in #✨｜sdxl

[untold glacier]

And the crop_w and _h?

[rough oxide]

haven't used those yet

[untold glacier]

Man I really need to find what I was reading yesterday.

[rough oxide]

but i think it's a literal crop of sorts, might work with the offset lora

lol

[rough oxide]

both are fast gens but sytan's is better organized and has better settings. that said.

if you want full generation, full steps

then you wanna look towards other workflows like my own

60 stepped dpm++ 2m generations will have better quality

than 20 stepped DDIM gens

finished testing a newer anime workflow too

[untold glacier]

I've found unipc_bh2 gives wonderful results with my little time spent with it.

[rough oxide]

kind of mixing my own w/ the new methods i learned xD

and then tweaking said methods even further

[untold glacier]

Nice!

[rough oxide]

and only going to 1280x1280 since i don't really have a need for 1536

uses more vram than any other setup i made tho since i use refiner too xD

major part of this one is using clip vit-l to guide the base image even further

should be more customizable prompt-wise than the other one

[untold glacier]

[rough oxide]

and i think introducing refiner prior to going into anime steps

kind of helps extra limbs a little bit more

but alas. bork hands. one day.

hopefully AI can solve that xD

[untold glacier]

Do you happen to know where to download custom nodes?

[rough oxide]

i do not but if you use comfy ui manager you can automatically dl custom nodes

i've only ever used the base nodes for making my stuff

if i went into upscaling i might look into them

[untold glacier]

This is starting to feel like

https://tenor.com/view/explain-plan-theory-crazy-insane-gif-22079506

Lol. But not worse than my Skyrim build.

[rough oxide]

a major reason for this upgrade (pardun the hands and messed up wine glasses cuz held objects still suck and these aren't final things ready for posting xD)

vit-l allows muuuuch more control over the environment/background

xD

than a pure vit-g fed into vit-l build

it can set the tones/themes better too

if you split the clips

anime model you had one job.

that job was to not turn feet into hands

LOL

[untold glacier]

What the heck is all over the sofa

[rough oxide]

wine stains

lol

idk about the last one

xD

[untold glacier]

Nah, that's like some strange mustard and alfredo sauce blend, lol.

It's like a cheese pizza exploded

[rough oxide]

finally a gud gen xD

[rough oxide]

gud enuff* lol

[zinc flax]

and then there's me with my simple casual workflow thingy

[rough oxide]

i still like to think my workflow is simple enuffff xD

one thing i always try to do just for my own use is put any box i need to put inputs in into the bounding box xD

[zinc flax]

It's actually pretty comfy once you get used to it lol

[rough oxide]

so it's there on an f5 or fresh entry

[zinc flax]

And definitely faster than a1111

[rough oxide]

yeah

[zinc flax]

But for stuff like inpainting... maybe not so much

[rough oxide]

actually*

[zinc flax]

Though I seldom use that stuff

[rough oxide]

well

do you know how to do inpainting in comfyui?

it's pretty decent tbh

[zinc flax]

Need to manually make the pic transparent iirc

[rough oxide]

nope

right click open in mask editor

#✨｜sdxl message

lol

my (i didn't know this comment right after)

xD

[zinc flax]

Ok now that's new to me lol

[rough oxide]

I AM TOTALLY AN EXPERT ON THIS

kekw

[zinc flax]

Back then there's nothing like that

Nice qol

[rough oxide]

niiiice

[wet elm]

i feel like almost everybody here is changing to comfyui lol

[untold glacier]

No longer than I have to.

[lyric zealot]

it's a interesting ui to play around, but when i just want to input some random prompt for generation, auto1111 is still what i go for

[untold glacier]

There's certainly no speedy way to make many small changes. But I haven't looked into custom nodes yet.

[wet elm]

the thing is that the only way i could maybe use SDXL is if i use comfyui but i dont want to have to learn it lol

[lyric zealot]

true

[untold glacier]

It's a lot to learn.

[lyric zealot]

just go grab someone's image that's generated with sdxl, drag it to your comfyui, then your node is all there

[wet elm]

its seems WAY to complicated

[lyric zealot]

just change prompt and you are good, sort of

[untold glacier]

#🍥｜anime message

[wet elm]

yeah damn

[wet elm]

and i dont have a lot of technical knowledge so idk

is there any chance that SDXL or auto1111 will have improvements to vram usage and speed?

[zinc flax]

You can also drag and drop gens from a111 to comfyui

Although you'd still need to play around with the nodes

[untold glacier]

What if I want to import the settings from a gen back into comfyui but without changing my entire node setup. Is that possible?

[zinc flax]

Not really possible since a1111 gen has its own "template" nodes on the metadata

[untold glacier]

[wet elm]

could anyone explain to me what is a "workflow", i see people saying this all the time here but i really dont know what it is lol (english isnt my main language btw)

[lyric zealot]

well, a workflow is like a step by step on how to reproduce the same pic

[zinc flax]

Not sure about this though since webui is pretty much a bit bloated with new features 🫠

[sage kiln]

Do you use ComfyUI or A1111?

[wet elm]

a1111, but im going to try comfyui today

[wet elm]

like a tutorial?

[zinc flax]

Workflow is basically the nodes on the ui

[sage kiln]

Yeah in A1111 the workflow is in your head. In ComfyUI you can arrange nodes that do the steps that you would do in A1111. This is considered a worklflow

[sage kiln]

That will really help him...

@wet elm just google comfyui workflow examples and start from the official github

[zinc flax]

#🍥｜anime message

like this

[wet elm]

ohh i kinda get it now ty

[wet elm]

before asking here i searched it but i didnt really undestand it

[wet elm]

ty

[untold glacier]

Yeah, it's the order and flow of nodes within the interface.

[sage kiln]

Btw I'm also new to comfyui. Is there a way to make a batch of say five images. Then I can choose the best out of those 5 and send it to an upscaler?

Or do I just have to use 2 workflows

[wet elm]

oh ty, and you can save them to use later?

[untold glacier]

You can, but they're also contained within the generated images themselves.

[wet elm]

so like metadata

[untold glacier]

So you can drag any of those images into the interface and it'll auto-load the workflow.

It'll completely replace what was there before though, so be careful.

[zinc flax]

Gets the job done for me
I don't really get the technical stuff so I make it simple lol

If I wanna play around with loras then I'll just add another nodes

[wet elm]

oohhhh that makes a lot more sense

[sage kiln]

Can you choose which initial image gets used for the hires or does it upscale all gens?

[zinc flax]

I actually haven't tried generating in batch so no clue about that

[wet elm]

i just launched comfyui for the first time. wtf.

[zinc flax]

Nice

Have fun lol

[untold glacier]

[gloomy scroll]

[wet elm]

how can i make the seed random in comfyui?

like -1 in auto1111

[untold glacier]

Within the sampler, change control after generate to randomize.

[wet elm]

ty

[wet elm]

bro added shading to the arrow LOL

[untold glacier]

Lol. It's an automatic feature of Greenshot.

[zinc flax]

I wish we can just use the previous seed instead of generating a new one

[wet elm]

could you share an image with me so i can get the workflow plz?