🍥｜anime

[dapper verge]

nai is what?

[rough oxide]

like if you do like

a prompt with 3-5 tags

or just

a 1-2 sentence prompt

nai might come out on top with style

but nai has that problem where as chars get further away

the eyes are really bad

a lot of models do

flux seems to be the first one where

that's not really the case a lot of times now

[glacial coral]

78 seconds compared to flux 300 ish second per gen i forget how quick any other model is

[rough oxide]

lol

for me something like nai would be 8 seconds

flux on standard 1024x1024 is

36 seconds

[glacial coral]

yh but ur on a desktop im on a macbook

i would use my 6600 xt with Zluda but im not in my room with my pc rn

[glacial coral]

what gpu u running tho

im still considering a refurb 3090 for £580 GBP

[rough oxide]

3090

[glacial coral]

Perfect at least ik roughly how long it would take if I got a 3090 🤣

I don't think £580 is a bad price either tbh

[dapper verge]

so nai is what? a model?

[glacial coral]

If u don't mind refurb or used I think it's the best under £600 option u could get for AI

[rough oxide]

it was the main anime model for the past like

loooong time

i feel like a lot of merges have surpassed it now tho

cuz it was like...

6+ months ago

[dapper verge]

ah

[glacial coral]

I would like to know how fast m4 pro is on SD

[rough oxide]

maybe 8-10 by now

still, if released it's massive.

bigger than anything else we have avail rn (aside from flux obvs)

but specialized in anime

it can even do some pretty obscure chars

[dapper verge]

have a link to the current one?

[rough oxide]

like lovely labrynth of the silver castle

it's a paid model

https://novelai.net/image

[dapper verge]

oh i see. so it's just a new version of it or a version of it for flux?

[rough oxide]

random flux gen for another thing i was testing for a 4ch post

naaaaah

if they're posting in the FOSS area of that server

and saying 5 days

either a. naidv3 gets FOSS'd

b. they release a new like naidv4

c. technical reports or sumthin dum (but would make future anime finetunes better)

something like that maybe

nai's main advantage

is it knows any anime char with any level of popularity

this is a random yugioh character for instance

that is just a monster on a card

and it gets the details near perfectly

cuz it's essentially

trained on all of danbooru

for

an extensive length of time

any char with like 80+ pics in the database would be usable

practically

even super niche stuff, and it gets all the details right

cuz the only origin of that char is just this xD

and it doesn't have nearly as much data on danbooru (at the time of training) as many other things

[rough oxide]

totes should've responded to this earlier, but ye, saaaame. AI is my absolute favorite thing, both text & image, sometimes music, primarily image lately... used to use stuff like TADNE/TWDNE models and other stylegan stuff before that, disco diffusion, etc... it's come so far in just 2 years, we can now get like exact anime screencap style stuff. rolling the gacha for a new anime image is just as fun as playing games xD

[glacial coral]

fr

[ivory pebble]

I'm guessing that message is about their new text model cause they've been working on that for a while now. I also can't seem to find that message on their server so Idk if that's even real 😭

[rough oxide]

nu cuz

it's specifically in a meme thread for naidv3 FOSS status

BUT IT MIGHT NOT BE MEME ANYMORE

is the hope

it's real

ik where it originated from, it's where 4chan gets a lotta info from

lol

but the question is... how real

[zinc flax]

inb4 it's just nai v4

[rough oxide]

i'd like... still be sort of ok with that

it'd suck but i'd go with it anyways

lmao

[ivory pebble]

They don't plan on finetuning flux so who knows what they're gonna do

[rough oxide]

we were all thinking wdv would be the thing to push anime gens into the future with finetunes/etc

[zinc flax]

This is the only realistic outcome lol

[rough oxide]

cuz of the sheer amount of tuning, being on par with or better than nai

[zinc flax]

Ain't no way nai v3 will go public

[rough oxide]

if naiv3 goes public it will be the best thing evr. that is all >:|

lol

it will seriously propel anime finetunes

into teh futureee

[zinc flax]

It's way too good of a model

[rough oxide]

not rly

it's been surpassed by flux and lots of model merge/finetunes imo

[zinc flax]

That's a take

[rough oxide]

the problem with naiv3 is u can't manually finetune or make loras/etc for it

as a base model it's good

but flux as a base model is even bettr imo

[ivory pebble]

It won't. They would be shooting themself in the foot by releasing that model. They know just how much money they're making with it, and making it open source is not a good idea for them

[zinc flax]

I find it hella impressive that ppl have been making flux loras already

How fast are they optimizing it lol

[zinc flax]

mhm

[ivory pebble]

Why not? It's an SDXL finetune, if it was open source I'm pretty sure it would be easy to make loras for it

[ivory pebble]

It's not a base model tho? 😭

[ivory pebble]

That would be amazing, but the question is, which base model are they gonna use

[zinc flax]

xl is the only option ig lol

Flux is also good but maybe next year who knows

[ivory pebble]

It could be their own base model, they did mention they wanted to make their own for a while now. I don't see why they would use XL again

Cause the only thing they could improve is the dataset by making it bigger and they could also add some tools like Controlnet or something like ADetailer for example

[ivory pebble]

Yeah flux's license seems to be the issue. Well not really, but for NAI it is

[rough oxide]

license is good for people who just use the model for outputs, but bad for if you're trying to profit off like specialized finetunes of the model.

my guess for nai thing is probably

both the simultaneous release of naidv4 and naidv3 getting foss'd?

pfffff. doesn't make sense to post in that thread otherwise

xD

[ivory pebble]

Highly doubt that, they didn't even want to release NAIv2 let alone v3, besides like I said it wouldn't be smart to do if they're looking to make money. But we shall see

[ivory pebble]

It's 4chan after all, they do a lil bit of trolling

[rough oxide]

no it's

literally tabloidA who posted it

from my own ss

[ivory pebble]

Ohh where is that from, which server?

[zinc flax]

nai v3 lite

[rough oxide]

ur in the server

[zinc flax]

featuring no artist styles

[rough oxide]

lots of stuff gets "leaked" there

before anywhere else

[ivory pebble]

Thought it was posted in the NAI server so I was looking there instead, my bad 😭

[rough oxide]

it wouldn't be a "leak" if it was on their own server. smh.

xD

ya tabloid said nothing in anlatan server

that i could find

[ivory pebble]

5 days is gonna be the new 2 weeks meme if nothing ends up happening lmao 😭

Lets hope for the best, NAIv3 on local would be insane

They better not get rid of the artist tags or else..

[zinc flax]

They already did with the furry model iirc

[ivory pebble]

Pretty sure they don't wanna risk getting a lawsuit from the western artists, I remember seeing something like that. They also don't plan on making furry models anymore so yeah

[bitter latch]

can't remember if I postedt his one but here ya go anyhow

[rough oxide]

has a whole pound of health for 4 mana, too op.

[bitter latch]

anyhow headed to bed in hopes of sleeping but knowing my body. be back in 3 or 4 hrs...sigh

[rough oxide]

gn

one day closer to naidv3 foss release.

[bitter latch]

FOSS , F-OVER Sus Service?

[rough oxide]

free and open source ;_;

wait

[bitter latch]

Found On Side Street

[rough oxide]

oh

i was missing a word

software.

free and open source software kekw

[bitter latch]

Zzzzzz

[unkempt cape]

To be honest I don't know lol

See the person

Is co owner of Anlatan so like unless it's a genuine leak

[bitter latch]

dreams

goes back to bed

[nova remnant]

Needs more bows

[ornate echo]

How about no.
Peddle your shit elsewhere.

[nova remnant]

thanos declines

Lmao

What was that about

[ornate echo]

Spammer trying to sell garbage.

[ornate echo]

These are great, btw. Nicely done.

[bitter latch]

thx

[nova remnant]

Giant girl

[ornate echo]

beeg

[green temple]

woow !

[bitter latch]

ooh

still not getting what i'm wanting.. dangit

ahh I see why. DOH

[slate narwhal]

Flux is lots and lots of fun. The more I use it the more I like it.

[bitter latch]

oops caught

adds random styles.. just .. cause

hmm supose to be a dragon...

AH HA!

[bitter latch]

changes main subject to bullet trains

[bitter latch]

ok not sure where the zombie cyborg came from

[bitter latch]

hmm has me wanting to reinstall cyber punk 2077

don't think I have the HD space

[slender barn]

last batch

[bitter latch]

interesting style

[slender barn]

thank you!!

[bitter latch]

there we go

ooooh

adds in style for tron legacy

"Pilot in position beginning startup sequence!"

I would so watch a tron anime

hmm which is better a lora, or an embedding?

there kinda more anime-ish

a stroll in the forbidden forest

there we go more anime-0ish

[slender barn]

Man these things get compressed.

[calm wave]

Ik

[bitter latch]

rolls a d20 to see if they stay up or not1-5 , bed, 6-10 stay up a little longer, 11 to 15 longer, 12- 20 I never sleep tonight

4. damn

2nd 20, tied. hmmm. best out of 3

4.. damn

guess i'm going to bed

best ouf of 5.. 15 and 2.. still lose

ok guess i'll be going now.. thud

bbl

[rough oxide]

playing around with sum tagging stuff i found on teh 4channels.

[nova remnant]

That's a lot of cat girl action

[bitter latch]

steampunk magical girl

[bitter latch]

o.. k then...

[green temple]

2nd ❤️

[bitter latch]

same prompt.. clipskip 10

clipskip 8

interesting effects

I should go back to bed.....

ooooh

ok last one ...

[dull glade]

Hi what model is this

[dull glade]

Hi what model is this

[rough oxide]

https://civitai.com/models/261336/animapencil-xl combined with https://github.com/KohakuBlueleaf/z-a1111-sd-webui-dtg

[dull glade]

Are you using automatic 111, sorry i am long lost to SD and not updated

[rough oxide]

for that specific model yes.

otherwise i normally use flux and that's comfyui

[dull glade]

which flux the latest ones?

are you using automatic 111 or other?

[rough oxide]

https://huggingface.co/black-forest-labs/FLUX.1-dev this is latest but automatic1111 doesn't support this model

u have to use comfyui or swarmui for it

but about to go to sleep for tonight, 4 am here, i'd have to give more detailed instructions later if you can't find em.

[dull glade]

i think i need a little bit help 😦

just downloaded the forge UI

[dull glade]

then how to install this?

[rough oxide]

animapencil would go in a models/stable-diffusion folder

idk how forge works

but i seriously gotta sleep sorreh ;_; it's way 2 late for me

[dull glade]

ok thank you

[fluid torrent]

finally managed to get flux running locally, with some decent speed

[nova remnant]

NF4 ?

[fluid torrent]

flux1-schnell-Q4_0.gguf

[nova remnant]

the schenell model isnt the better option

[bitter latch]

using model keyword and dantagGent ext

I like it

model keyword uses textual inversion and lora's. interesting

changes it up and experiments

[nova remnant]

some of your settings in that screen shot could be made to make your output better..... 1, change the shedulaer to simple not normal, 2. your CFG is at 15, bring that down to 4 or 5,,,... 3. put your steps up and around 20..... 4. lastly depending on your vram , change the image size to 576 x 1024

[vital raptor]

the flux schnell model is specifically made for low (4) steps, and the CFG is supposed to be 1

[nova remnant]

yes but that would be for cpu not gpu , those settings will still run it quicker, With a CFG of 15 thats where the vram killer lives, that needs to be 4 or 5 for optimal renders

and the flux model is trained with 576 x 1024

https://tensor.art/models/759856135286068673/FLUX-HYPER-TRAINED-DREAM-DIFFUSION-BY-DICE-V-1

I hyper trained the dev flux base

[vital raptor]

'optimal' is subjective, and model specific, but for gpu rendering the suggested settings are as I said, based on the official flux blog post

[nova remnant]

there wrong

that model is 60% under trained

[vital raptor]

according to what?

[nova remnant]

flux 1 and flux dev are also under trained but no where near as schenell

that model i just posted is the most trained version

[vital raptor]

I'm also curious where you found CFG somehow impacts vram usage, i've never seen that stated anywhere

[nova remnant]

and quicker than the others

higher cfg the more pull on resources

[vital raptor]

right, you said that, I'm asking for some kind of documentation on that

[nova remnant]

thats why sdxl ligtning and turbo models run with cfg of 1 nad 2 quick renders and less resourses

[vital raptor]

i mean, that's not why, it's the number of steps

[nova remnant]

if your a coder you would know this its simple fact

steps also pull resourses

steps add more detail, more detail means more ram

[fluid torrent]

my cfg is as low as 1.4 and its based on flux-schnell so it shouldnt need that many steps. ill try to upres a little bit, see where the limit is when it clogs up my system

[vital raptor]

no, more steps means more vram, each extra step (potentially) adds more detail

[nova remnant]

your t5 will also use more ram ie fp8 or fp16

[vital raptor]

the 'detail' of an image is not resource specific

[nova remnant]

omg

[fluid torrent]

schnell is just a filtered down version of the pro version, the huge flux model

[nova remnant]

im an ai developer for nearly 15 years and a profesional programmer for 20 plus years. i know what im talking about

[fluid torrent]

no, you are clearly not or you write it down the wrong way

[nova remnant]

yes noedel

[lyric zealot]

[nova remnant]

change the settings to what i suggested and see what happens bud

[bitter latch]

https://tenor.com/bY6rZ.gif

[vital raptor]

I'm just asking for documentation because everything you're saying is not what is officially suggested or what everyone else has documented

So either you're just spouting BS and using your model and 'proof' or you found something no one else has discovered

I'm inclined to believe the former tbh

[nova remnant]

take a look at my settings in that

and that prompt is the hardest to get correct till flux landed

[vital raptor]

i've seen something similar posted in the #sd3 channel

[nova remnant]

four bottles lined up on a table. from left to right, they are numbered "4" then "3" then "1" then "2". from left to right, they are red, blue, green, and orange. the background is a nightclub with a neon sign that says "Dream Diffusion".
Steps: 27, Sampler: Euler, Schedule type: Simple, CFG scale: 1, Distilled CFG Scale: 5, Seed: 1620423764, Size: 576x1024, Model hash: faf2118042, Model: Dream_Diff_Flux_V1, Version: f2.0.1v1.10.1-previous-297-g5fb67f49, Module 1: t5xxl_fp8_e4m3fn

the full set up

[vital raptor]

#🆕｜sd3 message

basically the same prompt on sd3

[nova remnant]

yes that was advertised by stabilty.ai as the prompt they said sd3 would do as no sdxl model could ever get that right

[vital raptor]

this is also the #🍥｜anime channel, colored bottles are trivial, show us some perfectly genned wifus

[nova remnant]

and i have run that prompt in all sd3 versions and it only got it corect twice

lol

[nova remnant]

spin me a prompt as i dont render that sort of stuff

and ill do it as we speak

[vital raptor]

I'm sure you can figure it out with all your extensive experience

[nova remnant]

i want that image you posted , post the settings here

we can compare that way like for like

[bitter latch]

you sure are asking alot... hmmm

[nova remnant]

lol

its only a prompt

[bitter latch]

to early for that

[nova remnant]

thats looks nice

[vital raptor]

solo, child, fox ears, long white hair, twintails, pink eyes, smile, hearts, close-up, hoodie, kneeling, outside

[nova remnant]

cool thanks ill test now and post the image

[nova remnant]

would you class it under anime, chibi, kawaii as i dont use these much

[vital raptor]

anime

[nova remnant]

ok its rendering 2 now

[fluid torrent]

thnx! i've adjusted my workflow

[nova remnant]

your welcome

[keen hollow]

I couldn't get results I liked in schnell... nf4 and dev working better for me, though obviously take significantly longer.

I guess there's an nf4 v2 out now, but not sure if it's any good or just sketchy junk.

[vital raptor]

flux is slow and needs some good finetuning before it's reasonable to use for most anime stuff

[nova remnant]

id pretty much say schenell is good if you run CPU rather than a GPU

[vital raptor]

prompt adherence is pretty much it's only benefit right now

[keen hollow]

random flux dev image

[nova remnant]

prompt adherence is pretty much it's only benefit right now...... and thats what makes it the best allround model

adherance is what every ai model has been ever trying to acheive

[vital raptor]

it's nowhere near the 'best' for making what i make, so i'll disagree with that

we're not here for the science, we're here for the cute anime girls

[nova remnant]

thats basically saying you cant prompt very well

[nova remnant]

as flux will with the right prompt output any style

[vital raptor]

and that's one hell of an assumption coming from some rando that joined today

get it to replicate my previous image style then

[nova remnant]

basically.... you could have just flux in your list of checkpoints and acheive any output you wanted

i did a few styles if you scroll up

but im simply only using your prompot you gave me

[vital raptor]

i'm not talking about a few styles, or 100 different styles, you said it could do any style, so i want my style

because if it can't do it, then it's not the best

[nova remnant]

which is a basic prompt doesnt even prompt the style of art in your prompt

solo, child, fox ears, long white hair, twintails, pink eyes, smile, hearts,, hoodie, kneeling, outside, hands on hips, tongue out

[vital raptor]

mrtarou is the artist from my prompt

[nova remnant]

the model knows it wants anime from the general wording

as some models will only render that prompt as realistic

solo, child, fox ears, long white hair, twintails, pink eyes, smile, hearts,, hoodie, kneeling, outside, but your prompt didnt state the artist mrtarou

[vital raptor]

because it wasn't important, and flux doesn't know what it means

[nova remnant]

thats the difference between a undertrained model and a correctly trained model

[vital raptor]

if you want the actual full prompt, here
mrtarou, solo, (child:1.2), petite, thin, fox ears, long white hair, twintails, (tilted head:1.2), pink eyes, score_9, score_8_up, score_7_up, source_anime, smug, close-up, hoodie,

[nova remnant]

the point is i didnt have the artist and generated the same style do you get what im saying?

[vital raptor]

no, the point is you said it can do any style, so i gave you a style to replicate

[nova remnant]

mrtarou, solo, (child:1.2), petite, thin, fox ears, long white hair, twintails, (tilted head:1.2), pink eyes, score_9, score_8_up, score_7_up, source_anime, smug, close-up, hoodie, thats apony model prompt

[vital raptor]

yes it is, good job

are you saying it can't replicate the style because it's pony?

[nova remnant]

well your using a bad version of pony if your output looks like that from your prompt. the pony checkpoint you are using is lame

[vital raptor]

oh, i didn't ask for your opinion, i asked for you to prove what you said

[nova remnant]

now go use your same prompt on my properly trained pony model https://tensor.art/models/751621497568661037/PONY-Dream-Diffusion-By-DICE-v1

[vital raptor]

I like what i get from this model with that prompt, flux can not replicate it

[nova remnant]

flux matched it with no required help. and your image wasnt the same propmt as you gave me

[vital raptor]

therefor, this is currently my 'best'

i'm not concerned about what someone else thinks is best unless it's functionally actually better

and your model with 2 downloads isn't likely to convince me, though the pony one might get closer than the flux one since it's pony based

[nova remnant]

you gave me this prompt solo, child, fox ears, long white hair, twintails, pink eyes, smile, hearts, close-up, hoodie, kneeling, outside but your render was made with this prompt mrtarou, solo, (child:1.2), petite, thin, fox ears, long white hair, twintails, (tilted head:1.2), pink eyes, score_9, score_8_up, score_7_up, source_anime, smug, close-up, hoodie,

2 completely different prompts

[vital raptor]

i'm aware

[nova remnant]

good bye

[vital raptor]

you didn't need the extra information to make a 'better image' in anime style, you just wanted a prompt

🤷‍♂️ i guess proving your claims is too hard

[steady grail]

too bad flux is too big for my 4gvram 😦

and only 7.2 gb of cpu 😦

i can't even load it on cpu

[vital raptor]

I feel my privilege saying it's slow on my 4090, but compared to sdxl/pony it's horribly slow

4-6x slower (on a 4090), mostly generic looking images (for anime) and the actual pixel for pixel quality (objective, not subjective) isn't really that great

lots of potential with it's superior prompt adherence, but it's not there yet and not even close to being the 'best' for anything people actually functionally use it for

[nova remnant]

lol i can see your GPU and its no RTX 4090. Who you trying to kid

[vital raptor]

you can?

[nova remnant]

and

[vital raptor]

omg, are you hacking me?

[nova remnant]

are you lying ?

4090 lol stop making out your something your not bud

[nova remnant]

its enbarrasing

[vital raptor]

keep coping man

[nova remnant]

where is the 4090

mine

now yours

im running 2 x RTX 4090s with 48 GB vram

[vital raptor]

and?

[nova remnant]

your full of bs

[fluid torrent]

hey dice, be nice

[nova remnant]

your not running a 4090, your running a potatoe pc

[nova remnant]

i am nice

[vital raptor]

whatever it takes to convince yourself

[nova remnant]

send the image i sent

[vital raptor]

i'm obviously not running windows you troll

but you can search this channel to find one i posted when i was

[nova remnant]

or ill reverse search that post and make you look even more redic

[fluid torrent]

that is not nice

[nova remnant]

oh so your running linux?

if your on linux its the same process to get to your dedicated ram info

[vital raptor]

it's not, but i already posted the screenshot

[nova remnant]

slide into your ubantu manager

[vital raptor]

ubantu isn't a thing, nor do i use it

[nova remnant]

i give up, crack on with your BS

[vital raptor]

be mad

[fluid torrent]

mr nice guy ...

[nova remnant]

https://ubuntu.com/download

not a thing lol

[vital raptor]

yes, that's ubuntu not ubantu

ffs man, lighten up

[nova remnant]

stop slating models people make for free for you to use and ill stop calling you out on your bs

[vital raptor]

i mean, i don't care

[nova remnant]

do you have any models you have made that are not merges?

[vital raptor]

not publicly released

[nova remnant]

lol ofc was expecting a answer along those lines

[vital raptor]

haven't needed to make my own since sdxl came out, it's a waste of time

[nova remnant]

so what model have you made that you resorted to using pony instead

you wouldnt know where to start with a check train

[vital raptor]

well i said i havent trained anything since before sdxl, so... what's your point?

[nova remnant]

https://www.shakker.ai/userpage/8b0d2aadaa2a4f2592cbb367c329ea51/publish my models 30 plus models

but your happy to use others work is the point

[vital raptor]

that's the point of them releasing it, yeah

[nova remnant]

and at the same time slating it

so your saying you trained a sd1.5 checkpoint?

[vital raptor]

i didn't "slate" anything, you're just getting your panties in a twist because i'm not falling for your crap

[nova remnant]

so your saying you trained a sd1.5 checkpoint?

[vital raptor]

yeah

[nova remnant]

ok 1 simple question for you

what did you use to train the check point and what coding lan did you use?

chop chop

[vital raptor]

i used kohya

[nova remnant]

fuck off

[vital raptor]

got the results i wanted, and that was it

[nova remnant]

laters

[vital raptor]

sorry, not playing your "i'm better than you" game, I don't care

[nova remnant]

you used kohya for a checkpoint.... bow your head in shame

you have no idea

[vital raptor]

you showed up out of nowhere telling everyone their images are bad and they need to use your settings and your models with nothing to back it up and then claim to be some amazing model developer and posting links to your model with 2 downloads, then start claiming I'm the one spouting BS?

[nova remnant]

you have obviously heard people saying training with kohya ..... koyha is for loras ya muppet

i never once said images are bad

not even in my nature

read up i was praising there renders

[lyric zealot]

[lyric zealot]

Ping me if there’s exciting new banter here

https://cdn.discordapp.com/attachments/1067638175478071307/1250543398591398001/GIF_2024-5-31_0-44-06.gif

[bitter latch]

hmm ok then

there we go

[fluid torrent]

(modern anime style, flat colors, lines:0.65), solo, sexy woman, fox ears, long white hair, twintails, pink eyes, smile, hearts, close-up, hoodie, kneeling, outside

[bitter latch]

this is with some other styles , cfg 3 and steps 10... using that ext dantaggen. wow

haha

that ext adds alot of details

[nova remnant]

I missed

Ugly gif

[fluid torrent]

flux

[twilit portal]

🙈

[bitter latch]

ok this can get random . wow

[nova remnant]

Are you talking about the image above or below

👍

[bitter latch]

ack

dammit

was the one withthe actual breakfast , didn't mean to delte

haha turned clipskip to 3, steps to 10 and cfg to 5..

[bitter latch]

little crispy

thats better

[lyric zealot]

https://media.discordapp.net/attachments/1270939143626686506/1270939144335659030/eliv_eating_plate.gif

[gleaming flame]

slightly pinup-ish... (made in invoke with counterfeit + atomix anime XL for upscaling)

[keen hollow]

I'm worried about their spines.

[gleaming flame]

these were made within leonardo ai, with their anime XL model... if anybody is curious

[gleaming flame]

I guess young anime ppl are quite bendy...

[nova remnant]

[keen hollow]

Neat. Do some widescreen ones. 😄

bonus points for 3840x1600

[gleaming flame]

more animeXL

[gleaming flame]

I will as soon as I get 4k screen... still working on fullHD...

[keen hollow]

your prompting looks interesting

[gleaming flame]

got a few "tricks" that work usually...

[keen hollow]

I've been messing around mostly with flux recently

[gleaming flame]

I heard it's quite nice. I wouldn't mind a update on resolution and IQ of my models, but with 6gb vram i'll stick to older ones. for now...

[keen hollow]

man I hate the time it takes to switch between models 😦

[gleaming flame]

That's why working with older models on my own pc and with new stuff online works best for me ;p

[nova remnant]

[green temple]

noice

[nova copper]

I have made a dataset for anime "screencap style" anyone can use without restraints (I want to see a flux lora finetune on this!)
AnimeScreenshots.zip

captions designed for flux

If you train a flux lora on this pls @ me :)

[green temple]

[bitter latch]

interesting 😄

[bitter latch]

with these other extension's going it's taking 5 mins per generation

but giving interesting results

changed settngs, not taking so long hmmms

[fluid torrent]

with a little help of flux, but with sd15 refine step

[nova remnant]

👍

Just in time when I needed a pillow

[green temple]

can`t see pillow

[bitter latch]

ok it went totally random.. wow

[gleaming flame]

elves... made with leo anime xl

[green temple]

this is better

[lyric zealot]

her during day vs her during night

[green temple]

oh , ok 😀

[lyric zealot]

actually they are probably differnet person cos of their eyes color

[green temple]

doesnt matter

[lyric zealot]

[green temple]

[steady grail]

i suppose that's one of the pleasures in knowing how to cook

[green temple]

yeah

[steady grail]

i would have prefered a guitar lol

i wonder if it gen a guitar in same pose lol

[nova remnant]

No

Contact lens

[lyric zealot]

stoopid

[nova remnant]

So many stopid gifs

[lyric zealot]

https://media.discordapp.net/attachments/1072697081443131476/1206242040052973598/0D682744-A633-4215-B958-55637CFEE0D3.gif

[bitter latch]

hmm must be a war vet

[dapper verge]

for leonardo anime xl, are you putting anything in the prompt to call out the sharp line anime style? I'm trying it out and it's on that model, but it's not looking like what you did.

[bitter latch]

now thats a franken truck

finally kinda anime japanese

wow

seems to like blond-ish hair

there we go thats better

so is anyone alive today? or just me?

[keen hollow]

I'm alive

[green temple]

me too

[bitter latch]

hope all are having a good day

[rough oxide]

just waiting for some news about anything. we haven't had a new anime model in ages ;_;

[nova remnant]

[dry knot]

give me a pic

[real imp]

https://files.projectlounge.pw/cta/files/ochonchon-20210517-0001.jpg

[dry knot]

/girl

[real imp]

You need to provide an image/GIF to add a caption! (Tip: try right-clicking/holding on a message and press Apps -> Select Image, then try again.)

[bitter latch]

ok that got random

1, 2.. wtf

o...k

hmm no better not

grr what is it with the closed eyes? oi

anyhooo....

nite

[gleaming flame]

I use lots of different descriptive words in my prompts, but simplest way probably: try adding colorpop, or soft pastel element (leonardo's own loras, 20-40% is usually enough). I just tried it on simple prompt, and here's the difference:

[bitter latch]

holographic colorpop

soft Line's , holographic colorpop, overlapping shades of light, blocky background

[nova remnant]

Angy

[nova remnant]

[bitter latch]

there we go

[gleaming flame]

wizardry school for girls

[bitter latch]

donuts!

whoa

[green temple]

fox girl ❤️

[dapper verge]

hmm, maybe i have to use the paid service, as this is what i'm getting with colorpop and the soft pastel element. maybe it would sharpen up with the quality setting turned on.

[dapper verge]

ok i feel like i'm getting somewhere with "Cute and simple anime flat illustration style." but i'm also finding it depends heavily on what's in the image.

[vital raptor]

Depending on the model training that’s fairly typical. Certain keywords can be over trained and will affect the image results

[vital raptor]

For example, “a girl” might produce a girl with short brown hair and brown eyes by default , but “a girl, blonde” will add blue eyes and larger breasts just because of the source material used

[bitter latch]

hmm that would explain why I get the same girl depending on the model. hmmms

tries a different tack.. racially diverse young woman

[vital raptor]

If you look at the last few images I’ve posted, there’s a pretty big quality difference, especially in the backgrounds, between the long white hair fox girl and the short blue haired elf

For some reason the combo of blue hair and pointy ears makes the model a lot more creative with colors, clothing designs, and backgrounds. The rest of the prompts are basically the same

[dapper verge]

yeah... i tried some anime stuff with their new phoenix model which is more prompt following, but it looks like it doesn't really know anime much at all. so i guess the ideal would be their anime model that's of phoenix quality.

[bitter latch]

ok that change does help

I need to think more outside whats normal to get different

[dapper verge]

leonardo / flux with flat color anime lora

[bitter latch]

simliar but different

[bitter latch]

there we go chibi

[bitter latch]

haha

[bitter latch]

hmm interesting effect

i did not prompt for speech bubble

I am loving the dantaggen Ext

[bitter latch]

what? I like my popsicle...

hmm can't seem to get her to be made of light

ponders

[vital raptor]

glowing skin?

[bitter latch]

like this

[green temple]

wierd but ok

[green temple]

😀

[nova remnant]

[bitter latch]

https://tenor.com/bw2Ly.gif

[nova remnant]

I was gonna gen smol

but got stuck on stupid ideas

[bitter latch]

done with work.. thud

brain.. tired

[green temple]

🌚

[green temple]

she is waiting smol

[nova remnant]

@green temple

I forgot again

👍

[bitter latch]

/me goes to bed early.. thud

z

[green temple]

wwwwpoooow

[stable crane]

I love the colors ❤️ ❤️

gorgeous gens

[bitter latch]

goes nuts , stackes ext styles and MORE

tries to break it

hmmm ok..

hmmmmm

oooh

random mode +9k

[bitter latch]

see...

I go random and it makes art....

puts them under water

hmm can't break it

ok fuk it, 51 card pickup!

[bitter latch]

the new matrix anime

ooh boy love this effect

[bitter latch]

clipskip 6 makes things wierd

but kinda nice too

clip skip 2

1 more and going back to bed, hopefully to sleep

clipskip 12

[slender barn]

Thank you 🙏🙏

[bitter latch]

yes she's covered

[nova remnant]

[bitter latch]

brings out , to fight it

[leaden stirrup]

naruto! #🍥｜anime

anime naruto

naruto #🍥｜anime

two girls in landspace! #🍥｜anime

[bitter latch]

there is not a bot here, stop

[leaden stirrup]

wow how can i use it

@bitter latch can you tech me how can i generate pectures

like yours

[brittle rampart]

good style.. but branches in hats....

[bitter latch]

I'm just using a bunch of styles with upscale prompt extenions 😄 , very random

but I like it

[brittle rampart]

well it doesn't look too bad...

[bitter latch]

then it does thjis. hhhaha

then it does things liket his... wow

[brittle rampart]

The style is reminiscent of old comics.

[bitter latch]

yep cause the first style I used was my retro anime new style 😄

changes models and vae

every once in a while it does eyes closed. odd

goes and turns up the cute factor

[bitter latch]

I took out all styles 😄 just to see what it would do

[bitter latch]

time for work.. sigh

[green temple]

wtf