🍥｜anime

[nova remnant]

very cute :p

[glacial bane]

That model really makes changing genders easy huh

[vital raptor]

and styles

it's pretty flexible

[glacial bane]

Leme squish it

[glacial bane]

Well have a goodnight I'ma sleep bye bye

Pass me that model ill check it out tomorrow appreciate it

[vital raptor]

Sorry, not public 😛

[glacial bane]

Understandable

[vital raptor]

😂

[vital raptor]

first try with this i tried "nice view" and it was just a butt shot 😄

[wraith raptor]

[nova remnant]

snow

moar snow

[vital raptor]

i thought i was gonna be clever and generate a picture of nothing but snow by using the prompt "snow", but i got this instead

[nova remnant]

Would pat this snow

[untold glacier]

Thanks. Yeah, I'm a fan of it. Coupled with BPN 1.3, it's hard to go wrong with generic prompts, let alone specific prompting.

Oops

Wrong link

https://civitai.com/models/148871/brightprotonukebpn-no-refiner-needed

Eface's model

[stable crane]

oh it's had updates?

[untold glacier]

The best model out there currently in regards to cohesion, and prompt responsiveness.

[stable crane]

will have to get latest

[untold glacier]

Get ittttttttt

Talking to Lewds

[stable crane]

I haven't diffused locally in a bit

I've been using a lot of NAI recently

you run the whole model on the phone? wtf

didn't know this was possible

wow impressive

what kinda speeds we talking?

truly

[nova remnant]

@wispy canopy you should try this model

I heard its good

[wispy canopy]

xD

[wispy canopy]

i heard the one who made it actually hates the model 😮

[untold glacier]

https://tenor.com/view/anime-laugh-lol-funny-lmao-gif-17671821

[wispy canopy]

yes way

total prick XD

who doesn't love his own child 😮

[nova remnant]

I said 'No Way', grabbed my bag and hit the highway ♫

reminded me of this cute and cool song

[wispy canopy]

it's the next thing i have akin to kids besides my cats

i love mine too

but bpn1.3? i love and hate it at the same time 😮

it's good, it's creative, and hasa few issues 😮

and i'm not sure how i can fix them

[nova remnant]

orange cats have 0.5 brain cells

Its common

[wispy canopy]

i have a calico -> so there's some orange in there

she's certainly the dumber one of the two

instead of opening a door with her paws (like my male tuxedo does), she just forces her face in between the opening up until it opens 😄

https://tenor.com/view/cat-squish-cat-cat-head-rub-head-rub-discord-gif-20128447

and of course -> you can't go to toilet unsupervised

[nova remnant]

facetank it

Softcore cat

[wispy canopy]

absolutely. anchoring me down by sitting in my pants xD

[keen hollow]

it was pretty simple... "anime style, video game boss, foliage, volumetric lighting"

evenin' folks

[keen hollow]

ugh I'm tired

[rich tartan]

Finally, got PLMS back.. Dam, this wasted a lot of time (GPU change

PLMS is still broken in the newer versions

[keen hollow]

try increasing steps a little bit

[supple raptor]

wait.... people still use PLMS?

[rich tartan]

I do, i love it

anyone got good negv prompt to fix those finger issues/recommendations?
My current one:

sampler

3D and ray traced look while keeping the anime design well

also the environment detail

(no loras, no extra detail adder)

Those fingers just issue a "little"

[supple raptor]

Less is more for negative prompts. A lot of what you have in there isn't understood by SD at all. It can't count so all of the three fingers, four fingers, etc stuff is just confusing it. Same goes for misplaced hand, misplaced foot and other tags like that. Images aren't tagged that way so it doesn't really mean anything. In Auto1111 emphasis between words is normalized. So when you have (1girl:1.2), red hair, what you're really prompting is (1girl:1.1), (red hair:0.9). I bring this up because all of those tags at 1.3 emphasis are making the rest of them less effective. This is all that I use for my negative prompt, adding to it only when there is a specific problem I need to counter with a specific prompt (such as the christmas tag adding santa hats to everyone): (Worst Quality, Low Quality:1.2), border, 2girls, nsfw, skimpy, grayscale, (watermark:1.2)

It's one of the oldest and least efficient sampling methods.

requires the most steps.

If you like the results you're getting with it, it's fine to use. I just haven't seen anyone using it other than as a meme to make cursed images at low steps in well over a year.

[rich tartan]

rly? im generating faster as dpm(any)/unipc

stopwatch says 2-4s

per img

rtx 4070

tho, it took a min or sth on my rx580 back then

[rich tartan]

is this better?

[supple raptor]

dpm++ samplers become coherent around 20 steps, DDIM is usually best between 30-40, euler/euler a is usually best around 30, and plms usually needs 50+. Outside of a few odd ones like DPM++2s a, DPM+ SDE, and Restart the it/s should be comparable between them with some small variances.

[supple raptor]

Yeah, that should be quite a bit better. I'm a bit of a disbeliever in the negative embeddings in general. After having used them for a year and then just stopping, the average quality of my images has increased noticeably. negative_hand does seem like one of the less harmful ones in my experience though.

[nova remnant]

[supple raptor]

cute

I don't even mind the claw hand

[rich tartan]

i just started using plms with 37Steps, earlier i used 25 (seems like old is gold):

but i should test the others one more ig

[rich tartan]

alr, i'll try out. Thank you

[nova remnant]

I have no idea what im doing I just wanted snow
1girl, (snow on the foreground, dynamic angle, fisheye lens:1.3), medium breasts, dress, playing on the snow, snowy, cute, smile, blue eyes, (high quality, best quality:1.3)
(worst quality, low quality:1.2), easynegative_8955, AuroraNegative, KHFB, negative_hand-neg_43127

[stable crane]

very crisp

[nova remnant]

no, you're very crisp!

Blazing drive model this time, not pikasnewgen for once

[nova remnant]

My sd is just really bad at bodies for some reason, have tried many stuff

[supple raptor]

honestly that hand isn't even too bad. it would be a relatively easy one to fix. the extra horizontal finger segment makes it seem worse than it is. extend that thumb a little, erase the extra segment, and draw in a knuckle for the ring finger and it would be very passable.

[nova remnant]

[native halo]

u missed all jojos i posted 🌚

[native halo]

ottoman empire general,solo,strong,macho

[autumn violet]

Why did I read that as oatmeal empire general

[native halo]

most history knowledgeable person in the us 🎓

[autumn violet]

Fear the oatmeal general

[untold glacier]

https://tenor.com/view/rtx-gpu-3080-nvidia-gif-20021847

...but generally, no. ComfyUI is already pretty well optimized by default and I don't know of any settings to change that would speed things up, other than not genning above the default SDXL resolution (no upres aside from the Ultimate SD Upscale tiling.)

Then make potato soup.

https://tenor.com/view/anime-food-anime-gif-anime-gif-gif-21989775

[terse ore]

Are there any dedicated SDXL models that do 2D style like Mistoon/Darksushi? All the stuff I see from SDXL is 3d looking anime or else 2d but from general purpose models

[untold glacier]

#🍥｜anime message

[native halo]

darksushi xl?

[terse ore]

Ah I'll check it out. I just used the search on civitai with "anime" and never saw it come up so yeah

[native halo]

its this one https://civitai.com/models/123008/dark-pizza-xl-origin-xl

[rough oxide]

nope, pure dall-e 3

[bitter latch]

oh this one's better

LCM

DPM++ 2M Karras in img2img

[vital raptor]

looks like the LCM one could use a little more CFG

[native halo]

it looks good if u have cataracts

[bitter latch]

changed a few settings and changed vae

lcm is a good starting point

[vital raptor]

different models behave differently with it as well

LCM works really well with my model

[bitter latch]

yeah there's a lot of different factors

[vital raptor]

based on what i've heard from other people, it works with sdxl model a lot better than 1.5

[bitter latch]

switched to my Fuzzy anime merge model and WOW

[native halo]

works with 1.5 too u just gotta play with settings i dont use it cuz i dont like the style

[bitter latch]

hehe the miku was to strong, it wanted blue hair, even tho I said red

I still like restart

and suppose to be bunny eared

eular A

[nova remnant]

That's not a good umbrella

Buy her another

[native halo]

i am not allowed to post the other pics i made of her

[nova remnant]

I bet the umbrella was too perfect for these jealous mods!

[native halo]

it was too windy so the umbrella went flyin

[cyan burrow]

almost holding it right, so close! cool pic though

[keen hollow]

evenin'

[keen hollow]

ok im done

ok now im done

[dawn jasper]

A Mortal Kombatant? 😮

[gusty palm]

cool style

[gusty palm]

make one where shes visiting f3mb0y hooters plz

[brittle rampart]

Hello guys! I came across this problem. After updating the extensions, the server stopped starting. the task just hangs on the command line without any signs of work or error messages. If I go through the execution, I get this on the command line...

i have a lot of disk space.

[vital raptor]

probably better off asking in #🤝｜tech-support and posting the entire error, the last few lines are more important than the first ones, but i'd recommend deleting the venv folder and then running it

[vital raptor]

my prompt was
cinematic anime style, woman with long hair wearing a long dress standing in a windy open field, star filed cloudy sky at (dusk:1.2), (from side:1.1), (expansive view:1.3),

i was trying to get more of a colorful sunset, but i like this one

[craggy turret]

What prompt do you guys think would result in this hairstyle

I've already tried messy hair

And weighting it differently like six times

[vital raptor]

you can try 'layered hair' and/or tousled

[vital raptor]

other words that might help are 'messy, wild, crazy, frizzy, bed head, edgy'

I tried mixing a few of them and get kind of close to the same idea, but i don't think my model knows that one

[craggy turret]

Hm, using the prompts,

(white hair, (tousled hair:1.2), (shoulder-length layered hair:1.3), wild messy hair)

I got this result

[brittle rampart]

It turned out that after the update Adetalier could not automatically download the mediapip

[magic wave]

Style mixing tests dump

[nova remnant]

https://tenor.com/view/cat-gif-20118453

[nova remnant]

Me on the left

[magic wave]

[native halo]

here is another on the left

[steady grail]

creepy

[cyan burrow]

lmao are we seriously gonna have to pay for commercial use of models released in the future?

so glad we already have sd 1.5 and sdxl and we can do just aobut whatever we can imagine with them

but apparently free until you make a certain amount? I hope that amount is large

cause paying a licensing fee for making maybe 100 dollars a month don't work 😄 Glad the older models are perpetually free and open

[bitter latch]

here ya go

[steady grail]

you can do better than this XD

[steady grail]

god, they are very sweet

i feel totally cold toward furry, i suppose

[bitter latch]

creepy anime doll

[steady grail]

yep

[bitter latch]

and then it does this

[steady grail]

wow

[bitter latch]

hmm no that one just creep "ME" out

Miku Mouse Mug shot photo , black and white (not what I meant)

cackles

country girl miku

[zinc flax]

Oh wow dollfie style

[untold glacier]

Reminds me of that one Fallout 4 mod, turning everyone's head into anime style heads, heh.

[nova remnant]

https://huggingface.co/playgroundai/playground-v2-1024px-aesthetic

@wispy canopy any thoughts of trying this out with 16gb ram?

[wispy canopy]

if you grab the safetensors edition, it behaves like a sdxl model

i didn't like it

it can't produce crappy images when i wwanted -> so it's biased towards an aethetic

[vital raptor]

based on their sample images, i'm not impressed tbh

[nova remnant]

i see

do you have some examples ?

[wispy canopy]

i'll have to make some

[nova remnant]

so this is their offical pic they posted

[wispy canopy]

@nova remnant currently i'm on team pixart
left is pixart, right is playground, just using some found song lyrics

[nova remnant]

pixart looks lot more interesting

[wispy canopy]

that's the thing

it's ALWAYS more interesting

i thought brightprotonuke1.3 was good

until i played with pixart

and ... sdxl just feels so flat and bad

[nova remnant]

i can see by example what you are saying

but as much i'd like to try it out ... i cant with a1111?

[wispy canopy]

when it comes to anime, it's not in a good spot yet

[wispy canopy]

i think vlad had some support for it already

but i don't know how well

[vital raptor]

it works

[wispy canopy]

playground or pixart?

or both?

[nova remnant]

or suppose i should wait some more days before its widely available

[vital raptor]

just use it like a normal sdxl checkpoint

playground

[wispy canopy]

quality felt very poor when i used it in a1111

[vital raptor]

haven't looked at pixart

[wispy canopy]

it's a beast to get running

mainly because of the t5 encoder

[vital raptor]

closeup realistic images look kind of amazing, but they're really closeup, and if you get any further away it's garbage

[wispy canopy]

if it's about pixart -> hard agree. but some dudes already working on making finetunes, and they're already looking 100x better

[vital raptor]

I think if you wanted to make cool "art" wallpapers with random stuff and no people in it (or people very far away) playground might be good

otherwise it can't do anime basically at all, and anything 'normal' and realistic is very meh quality

also no LCM

[nova remnant]

important take... i don't think i'd wanna download a 13gb file for that

[wispy canopy]

actually no -> lcm already built in, just only in their own backend

no one made a comfy node for it yet

or is it about playground?

[vital raptor]

i'm talking about playground

[wispy canopy]

i mean, the idea is there

[nova remnant]

i looked at some more examples with playground, not impressed by them

[wispy canopy]

i mean, for a base model, playground isn't THAT bad

[nova remnant]

yes, for a base model comparison that's understandable , but from user perspective quality factors gets a bit more demanding

[wispy canopy]

case in point about crappy bird, left is playground, right is pixart . I want CRAPPY

the part where i mention stick figure -> the bird ALWAYS gets a stick

no way to prevent it

[nova remnant]

well im gonna wait

but definitely things are cooking these couple of weeks

[wispy canopy]

yeah, better wait for some playground finetunes, if they are made

yeah, it's going at breakneck speeds again, like last year

[nova remnant]

exciting

[wispy canopy]

there's certainly potential in both models!

hoping we'll see finetunes of both

[cyan burrow]

I forgot how much I despise mechanical keyboards (when other people use them)

[rich tartan]

what models are those?

looks like sdxl

[wispy canopy]

first is playground v2

a sdxl sized model, but trained from the ground up

the other is pixart, new type of model, which relies on a HUGE tokenizer with actual transformer network for images, smaller than sd1.5

[rich tartan]

interesting

i may try out

[wispy canopy]

@rich tartan playgrounds has a sdxl compatible variant safetensors file, that one is plug and play

pixart -> radical different architecture, it's hard to get working on even comfyui

[rich tartan]

pixart? is that something like safetensors? like model.pixart

[wispy canopy]

no

it's completely different structure

[rich tartan]

aha, alr, thank you for the information, i'll look into that

[wispy canopy]

if you want to try : https://github.com/city96/ComfyUI_ExtraModels

their own implementation is ... crappy at best?

and with comfy you can offload the t5 encoder to cpu

[rich tartan]

t5 encoder? dam that's russian to me

Never used comfy tho

[wispy canopy]

i'd advice against using it for now

[rich tartan]

is it worth to switch over?

[wispy canopy]

for creativity -> yes

for ease of use -> god no

[rich tartan]

What is possible with comfyUI that's not possible with auto1111?

[wispy canopy]

the actual programming to run the model

a1111 is for clip only right now

which is drastically different from pixart, which uses T5 for tokenizing

or... whatever it does

[rich tartan]

did you do that with comfy?

[wispy canopy]

yes

[rich tartan]

hm, i shall try out new things too. But those nodes makes my eye hurt

[wispy canopy]

yeah, if you're new to it, you're in for a world of hurt

[nova remnant]

a1111 updates coming https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases

[rich tartan]

with and without(right side) xformers.. The eyes look better without(light effects)

[rich tartan]

ah nice, i also hope he/they also fix the broken samplers finally

im stuck at 1.5.2

their overwork on the samplers broke some of them

[vital raptor]

anything that increases speed (xformers) by reducing vram usage will cause quality loss, there's always a tradeoff

[rich tartan]

i think if i like one, i should just regen it without the optimizations. It may get better

[vital raptor]

as long as you also have enough system ram to compensate

the only time xformers is faster is when it reduces vram usage enough that it actually fits on the card, otherwise it's not a significant difference

but, at the same time, if you remove xformers for a high res gen it might get an OOM error

as an example, i can gen sdxl at 1024x1024 with a 2.8x hires scale 2867x2867 with 24GB vram and 64gb system ram (32gb shared) but if i go any higher i run out of memory, if I disable xformers, even the 2.8x scale doesn't work anymore

[supple raptor]

xformers gives both memory efficiency and acceleration. even if you wouldn't run out of vram with no optimization, xformers will still be considerably faster. if you meant compared to Scale Dot Product (SDP) optimization, then yeah the speed is basically the same but xformers is more memory efficient.

[near elbow]

minimalist anime

use another SDXL model but the original one - with the original, the additional refiner pass is needed for details which makes it awkward to use

"highly detailed intricate elaborate" should deliver more than enough

e.g. this is just "a cute anime neko" with SDXL anime style

same seed, only "highly detailed intricate elaborate" added - notice the difference? 🙂

well it looks more elaborate at first glance, also the belt buckle, flower in the hair etc

yozu can push it further e.g. with "((((highly detailed intricate elaborate))))" (still same seed)

or "((((highly detailed intricate elaborate fantastic arcane zany))))"

rest left as an execise to the reader 😉

change of scenery

[rich tartan]

Kitty came out unexcepted

[rich tartan]

Euler has some fog like overlay?

PVC

that one that i am using is not xl

dunno if that exists

[stable crane]

what's pixart?

oh a totally different architecture model?

[rich tartan]

yea

[cyan burrow]

+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

0

0

[supple raptor]

left, no contest. eyes and mouth are way better. an argument could be made for blending in the bangs from the right one into it though since the purple bits are a little cleaner in that one.

The star shaped pupils can look nice when they come out just right, but most of the time I think they just look messy. My model has a tendancy to do them unprompted which is one of the few things I don't like about it.

[terse ore]

@untold glacier I saw your anime comparison sheet and the pics on civitai. The Blue pencil model looks too versatile on civitai. It looks more like a generalist model. Did you have to go heavy on anime prompting to make it look more anime style? I don't usually have good luck with generalist models

[keen hollow]

evenin'

[untold glacier]

In the detailed comparison I did, all of the images were tested with the same prompts and seeds across the rows. This was to ensure that we'd be able to see the limits and features of each model properly, without tending towards any specific model's shortcomings or best qualities. The only models on that list that are not specifically made to do anime are BPN 1.2, and realistic freedom.

Many anime models that are made for anime can do realism rather well, as it's harder for a general model to do anime than it is for a anime-finetuned model to do realism. Generally though, if you want to try and make a general model to do anime properly, the prompt should include something like
flat colors, key art, 2d. (anime style:1.3) artwork of and then you put the rest of your prompt. You can also up the attention of the flat colors and 2d part. Also be sure to negative prompt:
(photo, photograph:1.3). 3d. along with anything else you don't want.
SDXL is highly flexible when it comes to what you're able to get out of a model, and if you prompt correctly, that'll get you most of the way there. General styles and gen cohesion with more difficult prompts are taken the extra mile with good models, especially ones that are fine tuned for what you're wanting to gen.

[stable crane]

neat concept

[untold glacier]

This... is the Stable Diffusion server, you know? 😝

[keen hollow]

LOL

ok ok ok

I'm actually running SD gens at the same time

lemme get agood one

[terse ore]

Yes, I figured that the prompts were the same across gens. I just wondered if the prompts were anime-heavy to begin with, which would somewhat negate the negative aspects of generalist models ie. Thank you for the tips though, I'll write it down. That must have been why I wasn't having much success.

[untold glacier]

I purposefully didn't put most of the keywords I usually use with general models for those tests; usually only including something like: "flat colors. key art" or something similar to influence the type of anime style I want for the gens. Every now and then I'd throw in a "anime style artwork of" instead before the prompt, if the rest of the prompt would try and veer away from that style by nature.

[untold glacier]

Nice! I like the flatness it has, and the colors.

Certainly the style of Alice I prefer.

[keen hollow]

I mean, I can do straight up Alice...

[untold glacier]

Nice!

[keen hollow]

gonna play with some sdxl stuff... is blue pencil still the best?

[native halo]

best at what

[keen hollow]

Dalle with the same prompt....

it's just ruined me.. lol

[untold glacier]

It was last I checked.

[keen hollow]

cool

i'll pull down the latest version

[untold glacier]

Well, aside from BPN. Often I start with that and then run BluePencil for the HiRes.

That gives a solid base for cohesion and then a great anime style after.

[keen hollow]

what's bpn?

[untold glacier]

BrightProtoNuke

https://civitai.com/models/148871/brightprotonukebpn-no-refiner-needed

[bitter latch]

broke people nanee?

😄

[keen hollow]

I see

[iron star]

this is one of my favorite anime of all time

[terse ore]

Do you upscale alone for images you like or you 1 shot the gen from start to finish? I must be doing upscaling wrong when I choose images to upscale separately because they always look like trash

[untold glacier]

Your denoise may be too high or too low, but I have to use ComfyUI for SDXL so it's a oneshot process for me, with multiple stages that it can save images at. Base gen, Upscale1, UltimateUpscale, BodyDetailer, FaceDetailer.

Also, never use the LCM sampler with upscaling, as it does terribly with it. Additionally, some lesser known models don't do well with upscaling.

[keen hollow]

it's an anime?

I frankly don't know anything about anime. 😛

[untold glacier]

Last I heard, it was a cartoon. A rather old one at this point.

[terse ore]

Yeah I use comfyUI as well. The problem is that 1 shotting takes longer because if the initial gen isn't what you're looking for then it still goes through the process. You could cancel it at the first stage I guess

[untold glacier]

Yeah, there's a lot of canceling I do.

C A N C E L C U L T U R E

[terse ore]

Do you also run it through an anime upscaler, like anime esrgan 4x?

[untold glacier]

There's two upscales in my workflow, so two of them. Both at 1.5x, which is the limit of current gen finetuned models that I've found. Ultrasharp for the first at a 0.5 denoise, and then the Ultimate SD Upscaler uses fatal_anime at 0.3 denoise.

If you're wanting a clean finish (less details where there isn't lineart), then animesharp or anime_6B is great for those.

[keen hollow]

Sadmose?

[untold glacier]

wut

[native halo]

https://tenor.com/view/the-what-gif-26849967

[keen hollow]

i need a good prompt to run in widescreen for a new wallpaper

[untold glacier]

Whatever you do, add "a stunning cinematic artwork" to the beginning and "by pascal blanche and peter balke" to the end.

[keen hollow]

it's already running a batch

you do one and share 😄

[untold glacier]

Oh I have tons of those

[keen hollow]

need one at 3840x1600

[untold glacier]

Well, sounds like you're gonna have to 2x HiRes then 1.5x with Ultimate SD Upscale, and hope that the first 2x HiRes remains cohesive. Depending on the model it's definitely possible, but one's pushing the limits at that point.

[keen hollow]

oh i do it all the time

I use tiled diffusion/tiled vae

[untold glacier]

Must be nice to have that with SDXL.

It's the best upscaling method by far. I wish that it worked with upscaling models when using SDXL models on Auto1111's build.

[keen hollow]

oh does it not?

i'm running oneo right now.. guess it might crash 😄

hmm it worked ok

i didn't set the cfg right so i gotta run it again though 😛

takes over 6 minutes

ugh

[untold glacier]

It didn't when I last tried it about a month ago. I had multiple problems that I had to solve, and that was the last of them that I couldn't fix. Also no half vae is the time killer on SDXL.

[untold glacier]

Nice

[keen hollow]

fixed the cfg and only took a couple minutes 😄

but i don't really want this as wallpaper.. need somethin' else

[untold glacier]

Yessssss

[untold glacier]

Try ultrasharp or fatal_anime as an upscaler though. You'll retain more detail. That and maybe up your denoise slightly on the last upres if you're doing more than one. That's my recommendation anyway.

[steady grail]

wow

[vital raptor]

15.2s gen time 😏

[steady grail]

wow

[steady grail]

at this point i will simply repeat myself lol

i honestly prefer generating square images and than cut thwm manually

[vital raptor]

holy...

[steady grail]

i love these gens, my god

[vital raptor]

thanks for that, this style is amazing

[untold glacier]

It really is, isn't it? 😄

Happened upon it by pure chance in my testing of/creating my Comfy workflow.

Well, partly pure chance. "stunning cinematic" took a lot of work in testing, figuring out what works best in prompting to give creativity but maintain cohesion.

[vital raptor]

😄

[untold glacier]

Every kid's imagination

[native halo]

average michael bay scene

[vital raptor]

lol

Transformers 12: Return of the Fallen Dark from the Empire of Beast Cannons

[untold glacier]

Don't think that's not where we're going when it comes to naming these things.

We're already getting close to Japanese LN levels with some of the stuff I've seen out there.

[vital raptor]

I watch enough isekai anime to know where it's going 😒

"I was just a normal salary worker and then I was reincarnated as a samurai toaster in a world full of cat girls"

[native halo]

That time i reincarnated as a girl panties

[primal wharf]

Regional prompter for chopsticks skinny legs

🥢

[vital raptor]

i should see about regional prompting again, last time i tried to mess with it it was broken in a1111

[untold glacier]

Likely still broken. The last time multidiffusion was updated was Nov 1st, according to its github.

[vital raptor]

when your city planners really like gothic architecture 😄

[untold glacier]

https://tenor.com/view/spike-trap-its-a-trap-surprise-pierced-fortnite-gif-14076664

[vital raptor]

nov 1st this year? because it's probably been 4+ months since i tried

[untold glacier]

When I tried it a month ago it was still broken

[vital raptor]

pretty sure it was before a1111 1.6.0 at the least

ah, that sucks

might be a good reason to play with comfy again

[primal wharf]

More chopsticks 🥢

[cyan burrow]

Hand!

[cyan burrow]

hello chaps

[steady grail]

ohy

[cyan burrow]

orly?

so...what are some must have games for the quest 3?

even through pcvr/steamvr

or on the quest store

[cyan burrow]

Obvious games though I suppose are the Resident Evil games that have VR support, then Skyrim, FO4 (vr versions, including the AI NPC mods), the Cyberpunk VR mod, Rogue Squadron, some other dog fight style game, some horror vr games that aren't about shooting things. Something that's like a first person rpg in vr. I don't want vr shooters yet, maybe on rails. I just don't want to vomit from moving around too fast

[vestal crypt]

absurdres, highres, ultra detailed, (red, black, white theme:1.3), landscape,
expressionless, looking at viewer,
(1girl, solo,:1.2),
gray hair, straight hair, parted bangs, (purple eyes:1.2),
Geisha Elegance, featuring graceful kimonos, elaborate hairstyles, and refined accessories.,
BREAK
Renaissance Chateau, renaissance architecture, frescoed ceilings, sculpted fountains, classical statuary, formal gardens, rich tapestries.,

[bitter latch]

Turtle girl

[cyan burrow]

wonder if SD 1.6 will come out open source or not

[bitter latch]

growing scales

[autumn violet]

Beatsaber, mod it with custom songs and play some of your favorite stuff. Unless your music taste is very obscure. That is like all I play on mine to be honest.

[cyan burrow]

yeaheventually, can't move around much now

and beatsaber requires some motion

[autumn violet]

True, vr in general can be like that

There are some things you can play sitting even though

I know tons of people ask for games like that, so if you search you can probably find threads with lists.

It's a bit hard for me with a lot of games since joystick locomotion still makes me want to puke

[cyan burrow]

nah not in general. Beatsaber wll need me to dodge stuff and swing around

my abdominal muscles haven't had enough rehab for that yet, neither has my diaphragm

[bitter latch]

ok back to bed.. thud

here have a miku doll

[rich tartan]

interesting widescreen, how'd you do that?

controlnet?

[native halo]

controlnet is not required for widescreen

[rich tartan]

woah interesting, do you need a specific model for that?

[nova remnant]

basic example you dont even need deep shrink with sdxl

[rich tartan]

ain't it make bad images

[native halo]

no,most models work with it

the ones that cant do it will show duplicates or deformities

[rich tartan]

🤔 im def gonna try that out, that looks cool

[rich tartan]

That's what i thought, what sampler you using for that?

[native halo]

depends on the model,some models work great with euler a,some with dpm 2m or others with unipc/dpm sde/restart

[nova remnant]

this one here, im usiong LCM sampler, LCM LoRA 8 steps and 2 cfg .. not the best pic but no deformities

[rich tartan]

hm, alright thank you for the information

[rich tartan]

LCM sampler? Is that one in 1.6.*?

[nova remnant]

LCM sampler and LCM LoRA both work on sd1x and sdxl

[nova remnant]

you will prolly need deep shrink with sd1.x

[supple raptor]

I got a graphic tablet today and drew for the first time in over a decade and my first time ever in anime style.

[native halo]

nope u can do it with normal high res fix

[supple raptor]

then threw it through controlnet

[nova remnant]

ahh damn, nice

[native halo]

we need a lora with this style

[supple raptor]

[nova remnant]

can you elaborate a bit more about the process you took?

[supple raptor]

just a simple miku prompt in txt2img hatsune miku, close-up, twintails, blue hair, blue eyes, parted lips, simple background, grey background

[nova remnant]

i didnt think it was that easy..............

[shadow smelt]

What are the ideal kohya hi res fix settings? I always get deformed bodies/artifacts

[nova remnant]

i was hoping for a long explanation about the cuteness you achieved in that one

[nova remnant]

this is no ordinary cuteness

[supple raptor]

I used my secret cute weapon ||Kakigori_v3||

[nova remnant]

ahhh i have to download that

thanks

btw how did you hide that text?

[supple raptor]

||text|| = ||text||

[nova remnant]

test here .. ||bla bla bla||

lol nice

[native halo]

||balls||

[nova remnant]

||big basket balls||

[rich tartan]

deep shrink, what does that do?

something like controlnet that generates areas?

[nova remnant]

like steven mentioned, it prevents duplicates and extra characters

[rich tartan]

alr thanks

[nova remnant]

btw .. LCM sampler and LCM LoRA with low steps and cfg is so good that i generated those images above at [00:16<00:00, 1.19it/s]

[rich tartan]

which card you use?

[nova remnant]

gfx? 3060 rtx

not just about card but with any 3x series you leverage on lcm tech

you can also get some great performance with 2x cards when using sd1.5

[rich tartan]

hm, im curious, does rtx 40XX have better performance in sd generally?

[native halo]

more cuda cores = more performance

[nova remnant]

well 4000 series are upper end, they should perform very well with sdxl

[rich tartan]

rtx 40XX has some ai improvements but idk if that applies to sd

[nova remnant]

i dunno about technical bits, but if you are planning on getting 4000 series card you won't regret it

[rich tartan]

i use rtx 4070

[nova remnant]

awesome

i couldn't afford more than rtx 3060 at the time i bought it

[rich tartan]

that's the cause why there is so much difference between my rx580 and rtx 4070, in gaming the 3x performance in sd (rocm and cuda) over 8x the performance

i thought it's just xformers or cuda

[nova remnant]

amd cards arent as optimized as nvidia cards

and you know nvidia even developed tensor rt recently

[native halo]

because amd doesnt have cuda cores and their ai cores suck

also tensor cores arent really that important for sd what matters most is cuda

[nova remnant]

||THE CUDA||

[rich tartan]

is that something similair to cuda just better?

[rich tartan]

yeah, that too

[nova remnant]

tesnor rt is not separate from cuda 🙂

i have to check the technical bit how tensor rt optimzes speed

[rich tartan]

is it already included in auto1111? Never head about it tho

[native halo]

but it uses tensor cores that are not as powerful as cuda

[nova remnant]

nice, that explains it

[nova remnant]

yes tensor rt is natively supported on SD

[rich tartan]

oh, that explains it

[native halo]

maybe they are more powerful but needs a software that can use them properly

[supple raptor]

TensorRT is nvidia's name for their flavor of ONNX runtime from what I understand. It's an alternative to Pytorch that's faster, but less convenient.

[nova remnant]

100% less convenient

wide screen closeup

[rich tartan]

i really would like to try the wide thing out, wanted to train my first lora and look into LCM sampler as i never heard of.. But my sd is on ubuntu and im downloading on windows..

can take a while

[rich tartan]

btw, does controlnet have something like api that allows external input?

I could use it with blender but was it controlnet?

[nova remnant]

wide res arent anything complicated, just a matter of specifying the size and maybe push it even farther with hires fix, you should be totally ok with sdxl models on your card

[rich tartan]

yeah, but i never used sdlx before, i have the base and i wanted to generate on v1.6.1 but there was some error thrown that sdlx couldn't be used or sth. Afterwards i downgraded to v1.5 because on 1.6 the PLMS sampler broke

[nova remnant]

well if you use automatic 1111 v.1.6.1 you should be good with sdxl but you wont need refiner for community developed sdxl models

[rich tartan]

it needs a refiner?

that's prob the issue

i didn't had one

[nova remnant]

the base model does but you dont have to bother with it for almost all the community models at civitai

[rich tartan]

alright, that's interesting, i'll retry then if im finally done downloading

[nova remnant]

but hold on, i suggest updating your a1111 first

[rich tartan]

on 1.7?

[nova remnant]

current version is 1.6.1

[rich tartan]

yeah, i'll do that, as far as i know 1.5 doesn't have sdxl support i think

[nova remnant]

yes you need to update

[rich tartan]

That looks awesome

what is the max height before you go out of vram?

[nova remnant]

not sure lol

2048x640

[native halo]

depends on driver version,some old versions allow u to go as high as u want without oom because it switches to system ram

[rich tartan]

Massive, dam that's good

[nova remnant]

something soothing

[rich tartan]

The thing is, rtx 3060 and rtx 4070 both have 12gb vram, so yea

[rich tartan]

It can generate really good environments

[nova remnant]

nice, and im NOT noticing any issue with 2048res so 4x cards would be even better

[rich tartan]

upscaling would lead to oom at that render size or?

and can you use upscaling with just denoising and steps but keeping it at x1?

[native halo]

depends a 2080 ti beats a 4060 8gb because of cuda cores

[rich tartan]

for more details

2080 ti is still a beast, ngl

[nova remnant]

i see

[rich tartan]

i may have some plans, can you use directml also with nvidia?

[rich tartan]

I generated those with driectml version on amdgpu, and i have never been able to reproduce such images on ubuntu. It could have great impact with that wide thing (but generation took like 3min or sth)

[rich tartan]

alright that's nice

[nova remnant]

hand picked 6 sdxl models

thinking of merging 1 and 4

[rich tartan]

you ever got good results by merging?

Mine only gets worse(noise adds in)

[nova remnant]

got interesting results

[vital raptor]

99% of the stuff i post is a merged model

[nova remnant]

i used the basic checkpoint merger on a1111 and they came out good

[rich tartan]

Interesting

then something's wrong with settings

[nova remnant]

lol you mean something's wrong with your settings

[vital raptor]

could be both 😏

[rich tartan]

ehem💀

[nova remnant]

:p

[rich tartan]

yes

[rich tartan]

Those are merged if i remember correctly, but you can see that there is much noise on them(right one was with plms and left one with euler a)

[vital raptor]

merging can take a lot of time and a bit of luck to get it right, overall i'm really happy with my merge, but there are a few things sometimes that just dont work right

[rich tartan]

you have certain things you look before merging? Im not experienced

[vital raptor]

i wouldnt worry too much about a tiny bit of noise at low resolutions

for me, I'm more concerned about the model producing people in a quality way, so I look for models that can give me a look but also have high quality eyes, hair, properly does poses, etc

[rich tartan]

Compare this and #🍥｜anime message with the merged ones

[vital raptor]

I don't post them here, but I do a lot of realistic images, so i look for that first

[native halo]

if u know what are u doing you can play around with weights when merging

[vital raptor]

generally I only merge something into the model if the one i'm using is missing something, like if i want it to be able to do realistic and anime, i find a good model for each that I like and put them together

[rich tartan]

dunno, i merged around on the model 6 times till the output got a little decent

[vital raptor]

most times it's probably better to just find a good model that does what you like and learn to prompt it properly

[rich tartan]

ah alr, thank you

[vital raptor]

prompts go a very long way to changing outputs

[native halo]

talking about this

[vital raptor]

block weights are confusing and i doubt anyone actually knows what they do

like most things, the math is magic and even the original devs are kind of guessing

but, using the supermerger extension in a1111 was really nice since you can set it to merge 2 models with random weights and see what it makes and do tests without actually saving a file

[native halo]

thats a cool opinion to have

[vital raptor]

makes it easy to try a random set, do some tests, randomize again and test more

[rich tartan]

that's actually something extremly usefull

[vital raptor]

that's how i did my merging for my sdxl model, I was also trying to preserve a bit of the lora i trained and merged into the base model (BPN in this case) but not preserve it 100%

[native halo]

yes it also has a preset thing so u can select any from list and see results

[vital raptor]

i ended up saving 3 actual versions of the merge over probably ~20-30 random weighted tests

then more testing, and landed on the current one

i've had a couple issues with it that have made me think maybe I should start over and try again so i can fix the problems, but then i generate something nice and don't really want to go through all the work again 😄

[nova remnant]

i take a simpler appoarch.. 🙂 i look for models that has good cover images, then download then and try out several different renders to get an idea what that model can do and after say 2 or 3 models that i like, i merge them in a % of multiplier and try them out a few different versions to see what i like

also when merging i focus on differences that i like to integrate

[native halo]

i take an even simpler approach,i wait for someone to waste time to make a good merge then i download it and use it

[nova remnant]

haha

[vital raptor]

i did that with my 1.5 model, started with a bunch of models i liked on civitai and just kind of threw them together with different % weights with the built-in merger until i found something good

[rich tartan]

lmao

[nova remnant]

i avoided the supermerger cause its too technical lol.. i should look into it seriously tho

[vital raptor]

yeah, i have no idea how it works, but the random weights part was basically needed in my case, doing a simple % split on the merge didnt give anything usable

so it was try and get a decent randomized merge, or start training with a new dataset and everything from scratch

[nova remnant]

yeah with a bit of random luck i managed to get some fairly ok modesl with merges 🙂

[vital raptor]

i wish i had kept better notes along the way though, i don't even know what's in this merge anymore

[nova remnant]

thats important to make notes

so i keep a simple note of the merge process so i can track back if things get too unexpcted

so these are some generic results of 2 different models with 3 in each set, curious to see what would result in merging those two models

[vital raptor]

yeah, one of the reasons I've debated starting over is so I can properly document the steps and what I used, then put it on civitai

I couldn't release this one if I wanted to just because I don't know who to credit 😄

[nova remnant]

lol

[vital raptor]

only model I know for sure is in here is BPN

i have a pretty short list of what it could be, but I did a lot of merging and testing with very little sleep for ~3-4 days

[nova remnant]

manged this generic output with those 2 merges

[vital raptor]

not bad

[nova remnant]

the eyes and hair leaning on the anime style that i wanted

think i will keep this merge with the rest

[nova remnant]

a valuable input came from you and steven about merging anime and realistic for the kind of sweet look i was asking about some days ago

this has nice consistency in the look, dont feel like tinkering with any further

[vital raptor]

one thing this model can't do reliably is better lighting 😩

[nova remnant]

i think multiplier 3 gives better look

has more anime feel

[vital raptor]

anime style, a girl 😄

[nova remnant]

what about cat ear

[vital raptor]

not there

one of the 'problems' with this model, it likes to put cats ears on the girls randomly

it must have been pretty heavily weighted somewhere in the merge

[nova remnant]

heh, thats a bit non generic imo

i came across some models that has fixed style in them

[vital raptor]

luckily it doesn't do it all the time, between these 2 i generated 10 images, 4 had cat ears

[nova remnant]

[untold glacier]

Lol

[untold glacier]

i n t e r n a l s c r e a m i n g

HOW DID YOU MAKE THIS!?

It's so good, even with the random protrusion on the front of the ship. I haven't had half a much luck on generating ST type ships in my testing thus far.

[steady grail]

damn, today i'm so much in reading Mode!

[nova remnant]

nothing too extraordinary really lol ....was just giving an example of 2048 res and made up a random prompt on sdxl ... something along this line i think

star trek, outer space, distant galaxy

[untold glacier]

Which model? And no LoRAs or anything?

[nova remnant]

just the raw prompt, for model i was experimenting with a merge i made

@native halo does this give you any error https://civitai.com/models/209030?modelVersionId=235424

oh btw, i used LCM LoRA and all the low settings

@untold glacier

[untold glacier]

Good grief that gif is much larger than it needs to be. snip

[nova remnant]

lol

[untold glacier]

This is quite the character

[nova remnant]

yeah thats weird lol ... i also had error popping up in my console

i wanted to cross check with steven since he might have used it or their other model

[steady grail]

what error do you get?

it may be a conflict with vae, or a lora

[nova remnant]

with BPN, full metadata in the image

[untold glacier]

Dunno. I've been seeing that myself. That's why people made CADS I suppose. I really wish that was available in Comfy, but it's in Auto1111 only thus far that I've seen.

Could be overtraining on certain things in some models.

The latent space is clearly pointing to similar things across the seeds.

[native halo]

yes i get tensor error

[nova remnant]

i got model load error after selecting chi chi

had to remove that one to fix the issue

[native halo]

yes both models give me that error

[nova remnant]

ouch

[untold glacier]

But remember, don't write a review unless it supports and loves him.

[nova remnant]

yeah thats whacked

[native halo]

Close the webUI and open the “config.json” file and delete the hash found in the setting “sd_checkpoint_hash” and delete the checkpoint name in “sd_model_checkpoint”. Then save the config and load up the webUI again, it should skip the initial model load from cache and work as intended, just need to select your desired checkpoint before generating.

thats the only way i could fix it

[nova remnant]

luckily didnt have too big of an issue, all i did was remove the model

[untold glacier]

Lol. Even in RC4 of Auto's 1.7 build, SDXL isn't viable. LCM and Kohya make no difference if it won't run properly to begin with.
A tensor with all NaNs was produced in VAE.

[native halo]

auto sucks for XL 😔

[untold glacier]

I love it to death, but it's growing more and more obsolete by the minute.

[native halo]

if comfy gets an ui like the auto one everyone would switch but i some guess ppl dont like the noodles

[untold glacier]

It just takes too darn long to effectively make big changes.

[untold glacier]

The only anime image I managed to gen before it said no more.

[sand geode]

anyone have some good workflows for comfyui? I am switching from auto1111

[nova remnant]

this is definitely the month of comfyui transition

never seen so many ppl switch in a month

[native halo]

dont know how many ppl switched since comfy or auto dont have an user count built in

[nova remnant]

try this https://openart.ai/workflows/dev

[sand geode]

thank you, civitai images clearly don't work for me so this is a lifesaver

[nova remnant]

im not switching to comfyui, if a1111 shuts down then i might

[sand geode]

I'm still using a1111, I am curious to see how this works tho

might switch

idk

[native halo]

if u use XL you may want to switch because performance is better at comfy

[nova remnant]

ok, i assume you are having specific issues with a1111?

[native halo]

you could also try with foocus

[sand geode]

nah, I was just curious really

[nova remnant]

ah ok 🙂

[sand geode]

i wanted to see what the "hype" was about

[native halo]

but for 1.5 the perfomance increase is not that big

[sand geode]

can't use XL lmao

[sand geode]

can't find any ones expressively for anime, any ideas where to look for images that give the workflow?

[nova remnant]

not really, i dont usually play with comfy

[sand geode]

tough

[native halo]

try to ask in #✨｜sdxl thats where most ppl post their workflows

[sand geode]

k thanks

[nova remnant]

i wonder why her face is a bit smudged

kind of like blury sketch

not very crisp

[sand geode]

I'm gonna not switch

I'll probably use both

I'm tryna get a novel cover

[nova remnant]

sdxl 512x768

didnt think 512x768 res on sdxl would come out this good

[stable crane]

amazing

[stable crane]

actually IDK if I believe this is AI now

[native halo]

human bean

[sand geode]

how did you get this?

[native halo]

from a real human bean artist

[sand geode]

Wow!11!1

[nova remnant]

1am Steven

[sand geode]

teach me your ways

[sand geode]

This is a1111, gonna look for a good workflow on comfyui tomorrow

[native halo]

my honest reaction

[sand geode]

https://tenor.com/view/star-wars-revenge-of-sith-anakin-vader-darth-vader-gif-19644107

[native halo]

https://tenor.com/view/not-from-a-jedi-palpatine-jedi-not-from-a-learn-this-power-gif-17298635

[sand geode]

https://tenor.com/view/darth-vader-im-ready-lightsaber-star-wars-gif-15851238

[nova remnant]

https://tenor.com/view/network-confused-john-travolta-wire-gif-5799547

Least complicated comfy UI

[untold glacier]

Now seems like a good time for this

[steady grail]

a close resemblance to my workflow lol

my workflow is just way bigger and confusing

[nova remnant]

[untold glacier]

https://tenor.com/view/revy-black-lagoon-anime-smiling-gif-22881392

[bitter latch]

out of the shadows comes the newist CyberMan

Miku Gets an upgrade

her mech

[bitter latch]

hmm some type of anime mech... I thnk

[vital raptor]

blue pencil is nice

[native halo]

thats a model for sure

[untold glacier]

As I said previously, I use NekoNSFW only for creative and aesthetic base level gens. It has major quality and cohesion issues that need an upscaler model with at least a 0.5 denoise to attempt to correct such issues. That and it really doesn't follow the prompt so well.

There's a reason why my previous comparison had Blue Pencil as number one in the list, heh. It's very good at what it does, if a bit lacking when it comes to specific flat styles of anime.

[bitter latch]

Gundam?

hmm or another anime.. looks familiar

[vital raptor]

looks more like evangelion

[bitter latch]

yeah

[native halo]

https://tenor.com/view/jake-peralta-squinting-read-trying-to-read-brooklyn-nine-nine-gif-13387080

[bitter latch]

hmm now I wanna watch a mech anime

[cyan burrow]

Miku

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1183329966050517072/00203-652506069.png?ex=6587f0e9&is=65757be9&hm=027766868fe809d7a2846f0c1e7607f250baf92cdcd75db30a480c1fbc80c242&=&format=webp&quality=lossless&width=441&height=662

[nova remnant]

and bloo

[hoary hedge]

Some prompts idea pls

[supple raptor]

my images basically always have PNG info

[hoary hedge]

So, something like a img2img? Or..

[native halo]

barack obama

[supple raptor]

You can run them through img2img or you can click the PNG Info tab in WebUI and drag the image there to get the full prompt and parameters of the generation like so:

From discord images, the easist way to get PNG info is to click the image, click view in browser, then drag the image to the PNG info tab in Web UI. If you use ComfyUI you can drag the image to your workspace instead and it will attempt to create a workflow for the image.

[hoary hedge]

Thnks, how can I repay you?

[supple raptor]

Just share nice images in here whenever you make them

[nova remnant]

Stay away from generating touhou

Or you'll end up like nubby

[supple raptor]

You say that like it's a problem

[steady grail]

wow. this looks an xl model

[native halo]

nope just good ol 1.5

[steady grail]

danbm what's that?

it's gorgeous

[native halo]

the kakarot 2.5d cozy one

[steady grail]

oh, that kakarot. still gives that weird error?

[native halo]

yea you have to switch models before closing webui otherwise it gives u that tensor error when u open it again

[steady grail]

i wonder if comfy gives the same error

i should try

i wonder if i can load two 1.5 models with 8g ram at the same time

a 2.8d? o.O'

[vivid karma]

How do you get such good details? 😮

[wooden coral]

Cool style

[nova remnant]

🐀

[near elbow]

wot?

yes YOU

don't anger her

too late

modernized

still not happy

[nova remnant]

hey @wispy canopy i rendered a bunch of 768x512 images with BPN and they all came out great, given that the model is SDXL

[wispy canopy]

😄

[nova remnant]

no freak deformity on single image

[wispy canopy]

still, i'm not happy with the model tho

[nova remnant]

why not

[wispy canopy]

it has issues

[nova remnant]

its doing its work

fix it in the next version 🙂

[wispy canopy]

i will ^^

[nova remnant]

i would hope you would keep the development going

and i find it useful that you kept the model versatile

i can just keep that model for general purpose

and no point in having multiple sdxl models

[wispy canopy]

^^ that's my whole philosophy of the model

[nova remnant]

just one request ... don't make the model render glossy crap 🙂

[wispy canopy]

yeah, it got a bit glossier, as one of the models used was closer to base SDXL

[nova remnant]

at present im happy with the model

[wispy canopy]

and had to increase the weightings on the OUT blocks to prevent other IN blocks going haywire

[nova remnant]

lol

im not dealing with supermerger so i wont understand the technical bits of it lol

[wispy canopy]

so right now i'm kinda stuck in a rough place -> i know some weights are out of order and will really go do weird stuff if make it worse, but i'm not sure how to repair that

i kinda developed a bit of a feel for it

[nova remnant]

well yes as you find the issues you can work on those things for better future versions

[wispy canopy]

knowing what knowledge is located in what blocks

but the last merger took a lot of effort, and left me a bit burned out 😦

[nova remnant]

ooh

dont burn yourself

might as well take healthy breaks and get better results

[wispy canopy]

of course ^^

[nova remnant]

i've tested several sdxl models so far

given a sdxl model does well what its supposed to do in versatility there is no point to use several sdxl models, they are not like sd1.5

so i decided to keep bpn the only sdxl model along with other sd1.5 for anime

also i think sdxl has lot to cover on anime aesthetics compared to sd1.5

[wispy canopy]

sdxl will prevail. king man will make sure of it! (this was on bpn1.3)

[nova remnant]

but the basic coherence with sdxl model is great and gives the scope to give good output

[nova remnant]

cool, but you see some of the sd1.5 anime faces that hasn't been achieved properly with sdxl yet

[wispy canopy]

yeah

bpn1.3 too realistic

can't go into that really cool expressions territory

[nova remnant]

it does anime fairly k too

[wispy canopy]

stuff like this, would be awesome if i could get it into 1.4... but this image is made on a 1.5 model