🍥｜anime

[bitter latch]

hmm prob all the artists confusing it

[steady grail]

the point is that without lcm, i get this

hard to say

[bitter latch]

fixes it

with euler

[steady grail]

lcm is getting influenced more by the model, than the prompt

[vital raptor]

you're changing the cfg and steps when you change to lcm right?

[steady grail]

yep

[bitter latch]

yes

[steady grail]

14 steps and cfg 1.4

[bitter latch]

50/5

😄

leaves 50/5 adds his hi-res mdjrny4 styles

waits

thats creepy in a wierd way

[steady grail]

yeah, it is

[bitter latch]

heavily edited prompt

[vital raptor]

The one on the left from the previous set is better i think too, but it's just the expression, so either could work

[bitter latch]

lcm doesn't always work for me

it's not very consistant

but can be a good starting point for img2img

[vital raptor]

I noticed some models are a bit more fiddley with LCM, but for the most part if you make small changes it's not hard to find the sweet spot

[bitter latch]

yeah and my merg of other merges prob isn't helping 😄

[vital raptor]

every 2-3 steps up or down can impact an image way more than it would without LCM

[bitter latch]

but I still like restart

seems my anime merge and my fuzzyanime merge are good so yeah

grinz evilly turns upto high on both steps and cfg with lcm... waits for horror

ooh no

[wispy canopy]

hope you like it ❤️

[vital raptor]

It's gorgeous

[bitter latch]

bpn?

[wispy canopy]

brightprotonuke

my checkpoint merge i released yesterday

feeling finally confindent enough it's decent

[vital raptor]

The eye clarity in closeups is really nice

for realistic gens

[bitter latch]

hmm sorta fuzzy around edges

[bitter latch]

that went in an odd direction

very sci-fi anime tho

[bitter latch]

fixed

[bitter latch]

hmm looks like a manga poster cover?

[bitter latch]

there we go

wait... the hands. wtf?

cropped looks better

ok thats it for me.. thud

[steady grail]

would have been so cute, without that disgusting expression XD

[valid turret]

Yeah, looks dumb 😣

[steady grail]

damn, that's the female version of the crow XD

[valid turret]

She is undead, look at that ear

[steady grail]

well, the crow, in the film, was pretty much an undead lol

[valid turret]

I have no idea about the film, sorry

[steady grail]

an old BUT good film of 90°

[valid turret]

ok thanks

[steady grail]

with a big story behind

2 chapter, should have been a trilogy

but do you know what happened?

the main actor got killed for real, in a scene

someone charge the gun, with real ammos

[valid turret]

that's horrible

[steady grail]

yep

the actor was very very good, that's why the impact was so much

never found out who was the killer

and the reasons behind that event

try to imagine a Leonardo Di Caprio dieing in such a way, that was the level of that actor

[wispy canopy]

❤️ very nice idea

[steady grail]

charming

oh, the actor of the film the crow, was the son of bruce lee... damn

Brandon Lee

[steady grail]

an hippie?

[wispy canopy]

found a setting for deepshrink that works, now to get the image interesting again

[steady grail]

guys, sdxl lcs generates better at low steps than 1.5?

[native halo]

better anatomy yes but it doesnt have all styles or anime characters that 1.5 has

[steady grail]

i was playing again with lcs, but with 1.5, i always get crappy results

[whole portal]

you mean LCM?

you are doing something wrong in that case, this is LCM (8 steps, WITHOUT upscale) #🏞｜general-with-images message

[steady grail]

yeah lcm

[whole portal]

show me your settings

and prompt

[steady grail]

finally, pushing cfg to 8, i finally get something remotelly resemple a result without lora and lcm

[whole portal]

and the output if possible

8? that i way too much

[steady grail]

i know, but look

without lora, i get this

with lora i get this

[whole portal]

you will never get the same image, even with identical prompt

please send settings for both

[steady grail]

with lora and cfg 8 i get this, ok, is crap, but it's way closer to the prompt

[whole portal]

do not use the same seed

[steady grail]

even with different seeds

without lora, i always get a big tree

with lora, i always get a big face

this makes no sense

[whole portal]

please send me the prompt

and settings

i am actually busy, but i will answer as soon as i can

[steady grail]

(masterpiece, best quality, highly detailed:1.2), made by miyazaki
2d game scene

HDR photo of landscape by Gustave Moreau,Thomas Kinkade,James Gurney. Carne Griffiths. Frank Frazetta. van gogh,Alberto Sevesooil paint,masterpiece,Realistic,deep colors,blue tint,only bronze gold moon,night scenery,fields,Field,Intricate,detailed,sharp,clear,Better image quality,harsh brush strokes,High dynamic range, vivid, rich details, clear shadows and highlights, realistic, intense, enhanced contrast, highly detailed

this is the prompt

(low quality, bad quality, worst quality:1.2), embedding:EasyNegative

[whole portal]

just copy paste the whole info

[nova remnant]

@steady grail please drag that image file in here and share

[steady grail]

the negative

[whole portal]

it is easier

[steady grail]

i don't know if it still have metadata

[nova remnant]

it does

just drag that actual image file here

[whole portal]

@nova remnant maybe meta is in file, but i do not have time to check 😦

sorry

[nova remnant]

i will check

[whole portal]

if you have time, please check

[steady grail]

let me look for the original if i still have it

[nova remnant]

or generate another one which gives you the big face

i am gonna load that up and see where its going wrong

[whole portal]

btw. there is a dirty trick i used to generate this #🏞｜general-with-images message

overweight some things that are not showing correctly

for example, my negative NSFW prompt has :2.0

the rest is normal

[nova remnant]

trick about what

i looked at the image

looks quite well composed image

[whole portal]

low cfg like 1.5, can ignore nsfw negative promt, but also some things in posiptive

use at least 1.7

and add more weight to things that "disapears"

do not increace cfg too much

2.5 is max ever

[nova remnant]

oh, i think i got nsfw even with it in negative when i used cfg 2

i tested that with eface's bpn

[whole portal]

try something like (nsfw, nude, naked:2.0)

and add the other things there

[nova remnant]

i see, might work

i only had nsfw in negative

[whole portal]

in positive use something like (pink hair:1.4) for example if hair is not pink enough

[nova remnant]

the prompt was .. a gothic girl in a tub of milk

[steady grail]

ok, generating something new for comparison, with metadata

[nova remnant]

and had nsfw in negative

this could be model dependent too

[whole portal]

cfg of 1.5 is too low for nsfw negative to be applied

adding :2.0 weight to negative prompt will help

[nova remnant]

what happens if you put sfw in positive and nsfw in negative

lol just thinking here

shaikorth you got one?

[steady grail]

generating

[nova remnant]

can you post the prompt ?

the text version 🙂

[steady grail]

it's the one up there

[nova remnant]

lol i dunno up where

if you can copy the text again please

[steady grail]

this

damn, freeu makes so much difference

[nova remnant]

nice ok

[steady grail]

without freeu i got this

[nova remnant]

so i got this

which model are you using?

[steady grail]

weird, the image is maked

[nova remnant]

and whats your lcm lora weight?

[steady grail]

soushiki

lora at 1

attempted at 0,6 but still got a face

[nova remnant]

and the sampling method?

[steady grail]

dpmpp_2m karras

without lcm

[nova remnant]

ok there is the problem... i suggest try the following for a test

[steady grail]

with lcm lcm sample and karras

[nova remnant]

sampling method DPM++ 2M SDE Karras
Lora wieght 0.3

[whole portal]

my advise is to use only one of those 3 with lcm lora - LCM, Euler A or UniPC

[steady grail]

i just noticed something, without lcm, during the gen, the noise get the form of a face, than get reworked and get the form of the tree

[nova remnant]

viking hang on

try what i suggested shaikorth

we can get you to have lcm sampler too but check that first

[whole portal]

unfortunately, some models do not work properly with any of 3 i mentioned ☠️

so, yes, try what torqx sugested

[nova remnant]

back on sd1.5 before lcm sampler was a thing i tested several images to come out good with dpm++ sampler like the one i mentioned, but had to reduce lora weight to 0.3

[steady grail]

something is cooking the images with the same set up 😦

[nova remnant]

lets hope its tasty

show the resutl please 🙂

[steady grail]

i experimented too much with my setting lol

[whole portal]

just do not use :0.3 with cfg 1.5 🤣

[nova remnant]

with all the metadate and settings

[whole portal]

i was so confused why i am getting totally washed up images

[nova remnant]

cfg 2

[whole portal]

almost forgot

[steady grail]

youtube was absorbing all my gpu damn hell

[nova remnant]

lol i leave youtube on most of the time or i get bored

[steady grail]

ok nuked out the workflow, let's work with your, mine is too messy

but i can't use xl

[whole portal]

slow graphics or you need 1.5 loras?

[steady grail]

slow graphic

[nova remnant]

can you load that image i posted?

[steady grail]

very slow

loaded it, but changed models

[nova remnant]

no need to change model

just apply the settings in image data and render a pic

[steady grail]

the model was an xl

[nova remnant]

wouldn't matter if xl or sd1.5

but try one that you can work with

[whole portal]

wait wait wait

xl models - XL version of LCM lora
1.5 models - 1.5 version of LCM lora

do not mix those

be sure that you are using the correct version of LCM lora for your model

1.5 lcm lora + sdxl model = https://www.youtube.com/watch?v=l482T0yNkeo

[nova remnant]

@steady grail here is the basic idea, given you are using sd1.5 now...

sampling method : DPM++ 2M SDE Karras
sampling steps: 8
cfg: 2
make sure to load the right LCM LoRA for sd1.5 (its different for SDXL so don't confuse the two), and use LoRA weight 0.3

[steady grail]

yeah i know

[nova remnant]

dont use hires fix atm, just check if your images are coming out right

[steady grail]

something is slowing down my gen by a lot

[nova remnant]

ow

not good

[steady grail]

yeah, weird

[nova remnant]

reboot the pc maybe and free up memory

[steady grail]

yeah, should

just gen the 2° img and than reboot

nothing, comfyui is frozen

[nova remnant]

can i ask, what's your vram and ram?

[steady grail]

damn, so boring

4g vram and 8 ram

[nova remnant]

ok, you shoulnd;t experiment with sdxl

sdxl requires at least 8gb vram

[steady grail]

for sure

XD

already did it, i can only do basic things

[nova remnant]

that would put heavy load on your system

not good to do that

can cause system crash

[steady grail]

yeah i know

[nova remnant]

use sd1.5 instead

not joking btw

[steady grail]

i did some risque things for curiosity

[nova remnant]

you can do all sorts of stuff with sd1.5 too

[steady grail]

even experimented with overclocking, not good things even when i knew what i was doing

i love pushing my babe to the limit

but no problem, i'm a tech XD

[nova remnant]

lol ok

but seriously over loading your system is not a good idea when you are running with 4gb vram and 8gb ram to run sdxl which requires lot more to function

[steady grail]

on paper, i'm an engeneeer in electonics and telecomunications... in reality i'm just a mad scientist <.<

i have loaded any sort of diagnostics studied too many papers for my own good xD

[nova remnant]

i see you are very curious about sdxl 🙂 upgrade the pc and you'll have lot of fun but to tell you, sd1.5 isn't bad .. just lacks some coherence but output can be pretty amazing

[steady grail]

yeah i know

i love comfy, because i can push gens to the limit

[nova remnant]

i couldn't settle with all the wire on comfy ui

totally confused me

[steady grail]

with this workflow i get these 2 results

the only difference are the lora, steps and cfg

and sampler

i try to generate something with only steps as difference

i'll try without changing steps too

already, i can see, the 2 results are similar

[lyric zealot]

god bless, anyone daring enough to update auto1111?

[steady grail]

finally comfy is generating as usual

i get anything, but not the tree

[nova remnant]

hey smol sus, whats the update about ?

[lyric zealot]

oh boy, it's a long one

[nova remnant]

released when?

[lyric zealot]

https://www.reddit.com/r/StableDiffusion/comments/18ah6u5/automatic1111_v170rc_published/

[nova remnant]

im curious cause i've heard big update was coming

[nova remnant]

nice, reading it now

[lyric zealot]

i guess it's not to main branch so

but it's release candidate, so it's close

[steady grail]

even with the same sampler, still not tree

[nova remnant]

ummm

is that recommended to update

also i wonder if they have official git hub page update

[lyric zealot]

i guess it's probably better to wait

[nova remnant]

yeah i wouldn't run with reddit

[lyric zealot]

yup here https://github.com/AUTOMATIC1111/stable-diffusion-webui/tree/release_candidate

it's the release candidate, so if you git pull, it's not gonna update

[nova remnant]

i see

[lyric zealot]

best thing to do is to wait for release candidate to merge with main

[steady grail]

looks like big stuff this update

but pretty itimidating

[nova remnant]

im looking at its featured updates

looks interesting

[lyric zealot]

yup, it's kinda scary

[nova remnant]

yes

[steady grail]

looks very radical

[lyric zealot]

Imma let early birds test the update

[steady grail]

oh, some nodes needed update

[lyric zealot]

[nova remnant]

@steady grail 4GB video card support (also reports of 2GB working) you might wanna know about that update

and @nova remnant

[steady grail]

oh, that's good

[lyric zealot]

https://www.reddit.com/r/StableDiffusion/comments/18agngo/fp8_support_added_to_the_automatic1111_web_ui/ i think it's this

[nova remnant]

Reloading checkpoints on the fly

No token limit for prompts (original stable diffusion lets you use up to 75 tokens)

[steady grail]

wow

being a comfy user, i can try to update

[nova remnant]

xformers, major speed increase for select cards

[steady grail]

it's not that i have anything to lose

[nova remnant]

smol sus the sweet bearer of good tidings

[steady grail]

i have it outdated just to experiment

[lyric zealot]

or you know, everything can just break apart

[nova remnant]

might initially but not when polished 🙂

im gonna hold for a bit before updating

[lyric zealot]

same

[steady grail]

it's good to know, too bad i'm already too into in comfy to return using auto1111 XD

[wispy canopy]

yeah, i saw that one on my civitai page, i think 😄

[nova remnant]

yeah was playing with it... it came out cute tho 🙂

[wispy canopy]

i think my lcm settings to go are lcm lora strength 0.8, lcm sampler, 12-20 steps, 2 cfg

before i had the lcm sampler -> lcm lora strength 0.3, dpm++ 2sa karras, discard final sigma (sampler setting in A1111)

did a huge comparison of settings, 2sa karras and LCM sampler simply work the best with lcm lora

[nova remnant]

i was doing something similar but after i had lcm sampler i use it most of the time with 0.6-0.8 weight

[steady grail]

why reducing the weight of the lora?

to reduce his sensibility?

[nova remnant]

well i found some of the render to have bit more contrast so i was playing with 0.6- 0.8

[wispy canopy]

it tends to overcook images on some models, i'm only using bpn1.3 so the sweet spot is around 0.6-0.8, as torqx also has found out

above 0.8 you start getting overcooking and and artifacts

[nova remnant]

yeah ... above 0.8 is too strong

[steady grail]

and here the return of the big tree

i love this model

[wispy canopy]

the dpm++ 2sa sampler doesn't know how to work with the lcm lora weights, and needs to only fed the lcm lora weights weakly

only dpm sde also got decent results

[steady grail]

i'll try that setting too

[wispy canopy]

the rest... simply is worse

[wispy canopy]

depending on prompt, there's artifacts, or it's actually really good

[steady grail]

close settings, but the first is withou lcm, the second, with it

[vital raptor]

@wispy canopy what do you use to merge checkpoints?

[steady grail]

the difference is radical

and without lcm i get a more close resemblance to the prompt

[nova remnant]

https://youtu.be/85r3pcKDvXI

[steady grail]

i should upscale that tree...

[wispy canopy]

A1111:

1. supermerger, add difference normal, to test random weights
2. supermerger, add difference train difference, to get a smooth model merge, without upsetting the balance of weights too much. This is after 10-20 random mixes in step 1, to decide what weights i'm going to use
   Comfy:
3. Block merging (SDXL transformers), again using a add difference method (have to construct the node order yourself) to fix any leftover weights you damaged in step 2

rinse and repeat

up until you have a merge that does what you want it to do

[steady grail]

maybe miyazaki in the prompt, makes things hard for lcm

being a very fantasy token

[wispy canopy]

lcm listens very well to whatever you feed it, in my experience

so, having conflicting tokens generally really upsets the prompting

[steady grail]

i need to experiment with a prompting very focused on lcm

[nova remnant]

that's one of the key aspect of LCM it would add bit more coherence to your sd1.5

[steady grail]

i suppose so

but i don't like coerence XD

i love suprises

[nova remnant]

you can aim for surprises with more generic prompts

and it would do that well

[wispy canopy]

i'm sure you'll love proper limbs too -> lcm fixes those very often

[vital raptor]

thanks, I'll have a look at supermerger

[steady grail]

oh good

[nova remnant]

you can narrow in on the specificaiton or make broad generalization and lcm would do that both with a nice touch in the output

[wispy canopy]

i'll be honest, i got used to lcm, and i don't think i'll switch over to another sampler back

restart was very good, albeit slow, but give the lcm lora some extra steps, and it's WAY faster than restart, with similar quality

[steady grail]

this is the workflow of someone that loves coerence? XD

[nova remnant]

looks dreadful to me ngl 🙂

[wispy canopy]

i'ma simple man -> i just type some text, pretty stuff fcomes out!

finger on her hand traveled ot other hand!

[steady grail]

you have no idea of the interactions between these nodes XD

i suppose i'll give lcm a chance

[wispy canopy]

and lcm is GREAT in combination with kohya's highres

[steady grail]

i still need to see a minimal influence of kohya in the end results

maybe it doesn't work with highres fix

[wispy canopy]

you mean, using kohya AND highresfix?

[dawn jasper]

This is an amazing image.

[steady grail]

yep

[wispy canopy]

you shoudn't need to

[steady grail]

yeah, i was thinking the same

[wispy canopy]

with kohya's you can go to highres immediatly

zelda is done with links bullshit

[steady grail]

yeah, as i was thinking

[wispy canopy]

you are thinking right, i can confirm this for you

[steady grail]

who needs bravery, when you have wisdom and a gun

[steady grail]

holy canuoli, that come out cool

[vital raptor]

So cute

[cyan burrow]

hmmm installed StyleTTS, gonna try and train some character voice from a game and see how well it replicates it

[valid turret]

Anyone here using the online webui provided by novita.ai?

[wooden coral]

Thats the right one

[valid turret]

IMO the best way to obtain creative prompts is to use an LLM model with an instruction for that

Or, wildcards

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1181278067461730385/00000-1272582296.png?ex=658079ee&is=656e04ee&hm=0a1548a3330e80f5c61561b54a48a3f1bdb0b4bc2013479e821b82754b93a192&=&format=webp&quality=lossless&width=441&height=662

[valid turret]

Nice one, too bad it's lowres

[rich tartan]

12gb vram good?

compared to 16?

[vestal crypt]

i think discord compressed the quality

[nova remnant]

@wispy canopy was SDXL Turbo part of the merge in BPN?

[wispy canopy]

[__medium*__|__medium*__] by ([__artists__:0.3]:1.3) , [anime line art:0.2], [{1-3$$__nouns__|__nouns__|__nouns__|__nouns__}:0.1], [[__movements__|__movements__]:0.1],desktop wallpaper , {abstract|surrealism|realistic} is my "i really don't know what i'm doing" wildcard

gives cool results

[wispy canopy]

no, i specifically left it out

so you can use it as a regular model too

turbo does degrade quality a bit, imho

[valid turret]

Nice one, mine is far more complicated 😄

[valid turret]

Wow, what is the tag for that hairstyle?

[gentle musk]

Mine?

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1181282482486788107/00024-2383934223.png?ex=65807e0a&is=656e090a&hm=acb927ace774fbcbb15082a56a9bd86f3bbe9c4f8b34de5196b3ed4cfdf7b6eb&=&format=webp&quality=lossless&width=441&height=662

[valid turret]

Yes

Wavy hair perhaps?

[gentle musk]

No specific tag, prompt was just "Anime Female Sherlock standing leaning against a Ferrari 360"

[wispy canopy]

added feminine in front

[valid turret]

So it's "victorian hairstyle", got it

[nova remnant]

i made a basic merge with 3 models among which one was turbo sdxl . and here is the result i got from it ...

[gentle musk]

Perhaps that might work

[wispy canopy]

not saying turbo is bad, but i notice backgrounds become a lot less detailed, as a general rule

[gentle musk]

It depends, the one I just posted is Turbo

[vital raptor]

between turbo and LCM on sdxl, my results were much closer to 'normal' when i used LCM, especially when I added in my custom lora i've been working on

[gentle musk]

But that's using the Turbo Lora, so you could just do the same with any other model.

[nova remnant]

a portrait of a beautiful woman with intricate jewelry and intricate floral background.

[vital raptor]

it's likely highly dependent on specific models and other loras used though

[gentle musk]

Turbo seems to be a lot better at lower steps

[wispy canopy]

except... the background is empty...

that's what i meant to say with it

[gentle musk]

But it has the added weirdness you have to work around if the images are more than 512x512

It can do more complex stuff with the right prompting

[nova remnant]

i see what you mean ... that's a basic portrait, but here is another example.. a woman in a lush tropical retreat with palm trees, sunshine, and ocean in the background.

[gentle musk]

I think the stuff it can do in a few seconds is pretty damn good though

[nova remnant]

the weight i used for merging turbo for it was 0.3

[gentle musk]

I've been using the LoRa around 0.4 - 0.6

Left is 0.3 and right is 0.55

[wispy canopy]

good stuff tho 🙂

[wispy canopy]

using the lora in a1111?

[gentle musk]

comfyui

[wispy canopy]

it gave me horrible results in a1111

[gentle musk]

Works brilliantly in Comfyui, a1111 is usually quite far behind

[nova remnant]

you have good experience on the merges, and im not sure why i miss a lot of the body below waist area to show properly with it

[gentle musk]

I think there is a new update today though, at least an RC anyway

[gentle musk]

Use full body at the end of your prompt, that usually works

[wispy canopy]

what are you promprint for?

[wispy canopy]

and get nudes: because "body"

[gentle musk]

Not had that issue myself

[wispy canopy]

had it often enough

but i never specify clothes either

[nova remnant]

fully body shot of a woman in a lush tropical retreat with palm trees, sunshine, and ocean in the background.

[wispy canopy]

you could try long shot, far away, that kind of stuff

[nova remnant]

long shot of a woman in a lush tropical retreat with palm trees, sunshine, and ocean in the background. she is wearing tank top and baggy pants.

i've bumped into this before with other prompts .. its possibly model specific to what i did with merges

[wispy canopy]

absolutely

[nova remnant]

i dunno what area to look for to fix that

but aside that i dont think quality is shabby

given i merged turbo

[nova remnant]

btw eface the reason im talking about turbo merge cause i was under the impression that could result in better coherence too

and some of the images on realism came out with beautful faces

something that surprised me

this was my merge workflow for it

[nova remnant]

a beautiful girl sits on the grass, her long hair flowing in the breeze. She is wearing a casual outfit. the sun is bright on a clear blue sky with puffy clouds.

semi real

[runic dome]

my ticket

[nova remnant]

i wonder if it's the booru tags or excessive use of comma separating every word gives the output an anime look w/o having to specify anime

[finite herald]

wallpaper desktop worthy wow

[valid turret]

I don't like the braided sidelock, looks too messy

[finite herald]

but that ponytail curl tho

[nova remnant]

bpn

a beautiful girl, in an enchanted forest, sits barefoot next to a clear water pond on a moonlit night.

[idle ether]

This belong here?

I don't post anything for almost a year

Time flies

[nova remnant]

dark winter night with a full moon, a girl roams in a forest with a torch in her hand.

[bitter latch]

Thank you

[dawn jasper]

Groovy image there too.

[nova remnant]

is there a discord server for Automatic 1111?

the rain falls heavy as a tiny frog finds shelter under a mushroom.

[nova remnant]

@nova remnant hewo

[granite void]

Girl

[nova remnant]

Girl indeed

girls make the world spin

deep in the jungle, close-up shot of a girl, sitting on the steps of a cozy hut, as snow falls under the moonlight.

[gritty spruce]

@nova remnant great lighting on all those last gens!

[nova remnant]

hey wyz thanks ~

also credit to eface for that model merge 🙂

[gritty spruce]

nice didn't know, is it on civit?

[nova remnant]

yep

https://civitai.com/models/148871/brightprotonukebpn-no-refiner-needed?modelVersionId=248861

[gritty spruce]

ty

[nova remnant]

🙂

[valid turret]

ShikiXL, it's in the filename

[terse ore]

Post full workflow

[supple raptor]

This just looks like it's being passed through the defective VAE twice. Download the latest version of the SDXL VAE from huggingface and it should solve the problem.

Oh, I didn't notice the sort of ghosting around the chin and cheek, there may also be a problem with the workflow causing the inpaint to not quite line up or something. The horizontal lines should be fixed with the VAE swap though.

[keen hollow]

what resolution and steps are you using?

ah, ok, you're good there

[supple raptor]

Yeah, 1 problem solved at least. The rest could be a model issue, prompt issue, or workflow issue. I'd recommend posting the full image directly from the folder so that people familiar with Comfy can take a look at the workflow to see if there are any problems with it. I only use Comfy a little and don't use adetailer in it.

[keen hollow]

ah, comfy

no degree in comfyology here

[native halo]

least buggy comfy workflow

[keen hollow]

try another image. try a closer portrait

[terse ore]

What are the main stuff people use in their workflow beyond the bare minimum? For me it's add detail Lora, face detailer, and anime upscale.

Oh, and negative embedding

[zinc flax]

https://civitai.com/models/132842/sweetparfait

I feel you lol

Okay maybe not to that extent LMAO

I see lametta

[steady grail]

most of the older models are obsolete and full of mutations

[native halo]

miku loves nvidia

[steady grail]

curse youuuuuuuuuuuuuuuuuuuuuuuuuu!

[native halo]

smallest 5090 variant

[keen hollow]

i like the hundos

[native halo]

prefer bananas tbh

[bitter latch]

that suppose to be a 4090.. doh!

[keen hollow]

this is a weird theme ya'll got

[native halo]

not weird at all i just wanna eat my gpu

[bitter latch]

yeah prob shouldn't do that oops

[keen hollow]

uh oh

what was wrong with it?

i just made some--no money thingies allowed?

[bitter latch]

this is better

[keen hollow]

Well.. I chuckled...

[bitter latch]

nice

[bitter latch]

😄

[bitter latch]

oooh damn

well that didn't come out as planned

[native halo]

she loves nvidia and money

she only has money and a pea brain so she bought a 6080 48gb card with only 128bit bus 😔

try to use high res fix ( idk how to do that on comfy)

[bitter latch]

LCM

nope

A merge of my own making 😄

Warcoal II : prompt: Miku Hatsune in a World of Warcraft Anime poster

yes

maybe I should publish this merge on civitai

ponders

[bitter latch]

replaced batman with naruto and went LCM

a new model that comes with animatediff

horror demon naruto

horror demon naruto

not gonna post anymore of those, they are getting.. shivers

ok gonna play some Starfield. nite

[hollow briar]

you can use postprocess to sharp the image,but the better way is to upscale the image and make in more steps

Well,maybe you can find some custom node to sharp

[hollow briar]

yep and you can try facedetailer too

[delicate vessel]

is there any self promotion channel in this server?

[gusty palm]

whats the verdict on these settings, mainly overlap

this is my img2img config

[nova remnant]

beauty

[gusty palm]

im trying to do something like thsi but with weed and hearts and geometry shapes instead of money, and also a blacklight colour scheme

and also on sd1.5

[nova remnant]

probably used controlnet

[nova remnant]

https://tenor.com/view/pat-pat-cat-pat-cat-cat-pat-love-cat-gif-8327343031414150876

not sure of the exact method but i think you can do that CN Ip adapter

[gusty palm]

this would be cool with controlnet canny

or lineart

[native halo]

no its just a handrawn style lora

[nova remnant]

@lyric zealot look at this noob

Not even 300 useless models that you don't even know what are anymore

[gusty palm]

my model folder looks similar, i wanna just delete a bunch of them but

[nova remnant]

there is some fun experiment you could try on img2img tab using controlnet

[nova remnant]

Bodysuit but mechanical

[native halo]

shes inside

[nova remnant]

lol ... uhh, thats fascinating

[nova remnant]

I once asked for a list of models smol was using

[lyric zealot]

[nova remnant]

Legend has it, smol sus is still screenshotting them to send to me

[nova remnant]

my guess.. she prolly has selected about 5 ish

[lyric zealot]

Oh yeah lmao

[nova remnant]

and best of the best

pikasnewgen best model

[lyric zealot]

It’s not that many tho

[nova remnant]

i had it bookmaked but not an option anymore since i moved to sdxl

[nova remnant]

Did not skip leg day

Now where is the girl

[native halo]

inside

[nova remnant]

(thick thighs, bodysuit, dark:2.2)

cute faces

[gusty palm]

something weird is happening

[steady grail]

weird but very creative

[wispy canopy]

ngl, i kinda like pixart, even if the quality isn't there yet

[steady grail]

refreshing

wow....

[wispy canopy]

perhaps someone recognizes the style

[steady grail]

yeah, the forms are a little irregular

oh god, gorgeous

[wispy canopy]

girl hiding her face behind a fan, a whimsical image of allure and intruigue, by yoshitaka amano

[steady grail]

let's try it out

oh right, driver updating

[wispy canopy]

yoshitataka amano being that dude that did the art for the early final fantasies

[steady grail]

ohh, good choice

[steady grail]

looks so much like a hand drawing

[wispy canopy]

that's tallgeese, isn't it?

i mean, i looks like a gundam series mech, for sure

but it gives me tallgeese vibes

[native halo]

almost all mechas from sd look like gundams prob because most of trainin imgs are from gundam

[steady grail]

or maybe because real robots are like gundam

forgot ip adapter on

🤣

but i like it

some tokens were werong, but holy moly

[steady grail]

far from perfect, but too tired to fix it lol

[toxic dragon]

Working on improving body shape and eyes while testing styles

[steady grail]

it's not connected, i'm doing fast gens

or do you mean theu are not sharp in general?

not what i wanted, but cute

soushiki

no

yep

but i suggest ft mse 840000

y

no idea

[steady grail]

this one

but it may be outdated

[steady grail]

just noticed some nodes were rendundant lol

lcm, euler, or dpmpp_2m or 3m

depend

most of the models are trained in euler

that when they were tweaking on the models, (don't know the details), they usually use euler as a base

so, the model respond better on euler

[steady grail]

even if it's very old

i agree

the details of the bg, the colors, are all fantastic

and respond very well on any lora i used till now

my workflow is a true frankenstine

i generate something with metadata, but it's still a work in progress, i'm fixing it

instead of nuking everything, i'm just adding nodes on nodes and never touch what works even if i downloaded better dones

so i have this mix of old obsolete nodes, mixed with new nodes

some were even connected in loop

i must have connected something wrong...

nothing, frozen

wow

cute face

but the shadows are weird

adetailer acting up?

wow

damn

i broke something....

yep

i suppose i should nuke everything and remake the entire workflow from 0

this thing is too much intricate

i know, right? XD

top right is ip adapter

in the middle, i have high res fix, than adetailer and in the end, upscaler

yup

assembled piece by piece

but now it's too intricate and i need to purge it

under the nodes, there are other nodes you can't see

try these settings for adetailer

should fix the problem of the shadows on the face

connected this node istead of yours

do you need the full adetailer?

k

i should go too

both cute

hard to chose

[hollow briar]

More steps maybe will help

[steady grail]

i would delete that

is definitely nsfw

charming, reminds me of re zero

[hollow briar]

Ok I will make more cloth coveing

[steady grail]

obscene images are against the rules in this server

when a mod comes, you will get kicked out

so, plz delete it

[hollow briar]

Sorry for that, deleted

[native halo]

because sex is popular

[steady grail]

i suppose so

i know just filter it out

even nudity is necessary for a stable model

without nudity, you will get foul anatomies

[wooden coral]

[steady grail]

yep

eheheh

[native halo]

cant wait for sex 2 official trailer

[steady grail]

ehehe

i learned anathomy from leonardo davinci sketchs

that guy was a genius

really mind blowing

[wooden coral]

For that youd have to watch the first one

[toxic dragon]

Brain rot immediately turned my mind to fgo when I heard Da Vinci. The mind virus really is taking over.

[steady grail]

fgo?

fate?

[toxic dragon]

yes

Privately involved in a campaign based around that with custom servants. 🙂

[steady grail]

i'm totally in love with the art and life of da vinci

my love for art is born with him

[toxic dragon]

same

[steady grail]

art is expression

just find the form better attuned with your soul to express your feelings

people can't do what they are not attuned with

they can learn the technique, but never make that art theirs

there an absolute difference from imitation and inspiration

nighty night

[toxic dragon]

bye

[steady grail]

JOJO!

[gusty palm]

yall up making arts at 5am

[toxic dragon]

Finally finding my footing again.

Although I need to download that new inpainting thing I got sent a while back lol

[lyric zealot]

it's 5am for you, but 6pm for me

[gusty palm]

baw thats so cool, i wish i lived somewhere else

me irl

[gusty palm]

i wanna sell me stuff as stoner posters but i feel like people "ick" at the thought of ai art

even though a majority of stoner posters are highly generic lmao

[lyric zealot]

[nova remnant]

would pat

https://media.discordapp.net/attachments/493156996644864001/1140621716356337674/CATOJERK.gif

[toxic dragon]

This came out a little sketchy but oh well.

[native halo]

i love her big eyes

[toxic dragon]

The eyes are a big improvement over what I did a week ago :))))

The only thing I'm messing with rn is making the face looking more mature

[nova remnant]

what is this about ... i downloaded that from huggingface page

*** Error verifying pickled file from C:\StableDiffusion\stable-diffusion-webui\extensions\sd-webui-animatediff\model\mm_sdxl_v10_beta.ckpt
*** The file may be malicious, so the program is not going to read it.
*** You can skip this check with --disable-safe-unpickle commandline argument.

[cyan burrow]

hmmmmm

am I connected?

[bitter latch]

nope

😄

[cyan burrow]

I figured

I've discovered that I really like darksynth/dark synthwave or Dark Cyberpunk music

as it might also be called

like Biowave by Tokyo Rose, Comeback by Lazerpunk (doesn't get started until like a minute in so patience), 777 by Draven

[wooden coral]

Dark Synthwave is a great genre

[dawn jasper]

This woman and her measurements are lovely.

[wooden coral]

anyone here managed to generate a waterfall out of the mouth?
Im trying right now and it feels impossible

You either get a waterfall or someone with an open mouth

[native halo]

yes just photobash the waterfall then img2img 🧠

[wooden coral]

I mean yeah, but trying to generate it from scratch rn

[native halo]

i tried to generate the tiny miku vs giant gpu but couldnt i guess there are some things sd struggles to understand

[wooden coral]

Frkn nightmare fuel

Also batch size seems broken in A1111, generates one image with the prompt and the other 3 without it seems

Setting batch count to 4 works tho

[cyan burrow]

why are sometimes images blurred on dreamstudio.ai ? It's not that they were very inappropriate and removed entirely, they appear as blurred and I can't remove the blur

but I can clearly see behind the blur that there's no nudity

I mean I get this crap

[wooden coral]

Maybe it gets falsly flagged?

The bot here does that too sometimes

[cyan burrow]

but then it would be removed

not blurred

They still charge credits for blurred but I can never unblur them

[native halo]

dalle can get waterfall but it always put a tiny miku inside mouth

[wooden coral]

It write a support email if I were you, cause removing credits but getting no output sounds kinda sus

Oh right I can try dalle

[cyan burrow]

I don't want to pay for that. I won myy dreamstudio credits when CoW was around

[wooden coral]

I mean you dont have to pay to get support? At least id think so

[cyan burrow]

now it works better

[wooden coral]

non anime

Dalle is so good at prompt understanding

[cyan burrow]

time to play cyberpunk

if I can get it to work with any of the mods I have

[wooden coral]

Have fun, I enjoyed my 80 hours in 2.0

Dont see a reason to play 2.1 now unfortunately

[native halo]

[nova remnant]

which one do you like better in comparison, they used two different vae

actually let me use same seed for better visual

[wooden coral]

I like the second one more, looks more saturated

[nova remnant]

me too

[native halo]

i like the one with broken hands

[nova remnant]

that was with https://civitai.com/models/152040/xlvaec altho the author puts that into ongoing development

shared by @whole portal 🙂

this sdxl vae variant is like sd1.5 anime vae

got this with that new vae while playing with animatediff

[wooden coral]

Gifs just suck color wise

[nova remnant]

lol yes

the mp4 didnt come out too clear either

i think it could be the settings

[wooden coral]

Yeah image clarity is another thing, but gifs are reduced to 256 colors I think

Which aint a lot

[nova remnant]

needs more inception

exactly

gif is not an option

but i went with it ... first time trying out animatediff

[wooden coral]

Only advantage is size and it loops

Try the VHS video combine node

If you arent already

[nova remnant]

i used ffmpeg

but thats for mp4

[wooden coral]

Are you using comfy?

[nova remnant]

using a1111

[wooden coral]

Ah well

[wooden coral]

If fruit would see what you do with miku sometimes

https://tenor.com/view/spongebob-ban-pubg-lite-banned-rainbow-gif-16212382

[nova remnant]

steven loves miku .. has a secret crush

i love this vae

the whole image composition looks so much brighter

[nova remnant]

a girl in a futuristic cyberpunk setting, with neon-lit skyscrapers towering over her, stands on a street adorned with holographic signs.

my checkpoint merge is broken

it goes all the way and fails on saving the newly merged file

TypeError: argument 'metadata': 'dict' object cannot be converted to 'PyString'

[vital raptor]

are you using the a1111 merger tab?

[nova remnant]

yes

[vital raptor]

if you open the metadata part at the bottom and uncheck 'save metadata' it'll save

i don't know how to actually fix the metadata though

[nova remnant]

let me try that suggestion of yours

merging now

i looked up that bug, they dont have any fix for it yet

but if your suggestion works im happy ... lets see if it can save

awesome it worked thanks 🙂

[nova remnant]

fingers came out really well for picking guitar

[wooden coral]

Either I suck at prompting or dalle just is that good

[steady grail]

i would support that decision XD

[nova remnant]

looks great

dalle interprets prompts really well

but i still like SD image textures better

dalle can't make cute faces like SD

[steady grail]

his gens achieves a new level of curse XD

[nova remnant]

cute faces like these are heart pleasers

[wooden coral]

Generic anime face

[nova remnant]

but sweet

one of the reasons i was hesitant to play with sdxl before is when most trained sdxl models initially didn't come close to sd1.5 cute faces

cowboy shot gave her the cowboy hat ...
beautiful girl, dark hair, gray eyes, open jacket, shorts, outdoor, bokeh, cowboy shot

hat in negative .. easy quick fix

the next a1111 update coming out soon, will address those prompt coherence issues

[lyric zealot]

[nova remnant]

squish

[lyric zealot]

@wooden coral let me know if it's considered nsfw

[nova remnant]

nsfw = not working at office desk

its that weird lol

[lyric zealot]

[nova remnant]

i would embrace reality and life's natural beauty

[nova remnant]

the implication of suppressing feminine beauty can result in negative energy

this merge came out interesting

[nova remnant]

im kinda going crazy with booru tags on sdxl lol i think it lets the ai decide for itself what to do with all that word options

[nova remnant]

i deliberately prompted this
pretty face, girl, long bob cut, red hair, tank top, skirt, huts, plants, flowers, outdoors, lampposts, pond, butterfly, rock, puffy clouds, earrings, necklace, sitting, sneakers, evening,

this is the result

[lyric zealot]

V?

[nova remnant]

@wispy canopy i got something intersting by merging your model

[wispy canopy]

what did you do? ^^

made it an anime model?

[nova remnant]

not entirely anime, its still got the versatility but i merged it with another versatile model and the results are pretty sweet