>no forge neo in op
anti-gradiomitic thread

File: 00755.jpg (871.8 KB)

871.8 KB JPG

>Why are you doing [thing]
Why not, I am just that bored so anything goes I guess
>Why is so [thing]
Because it's me
>Are you going to do [thing]
Maybe I will maybe I won't
>Can you [thing] more [thing]
No, usually I determine the degree of [thing] while working on the gen with no plans on adding more of [thing]

>>8834453
Oh yeah, we should start adding that one for the next threads

>>8834452
Style wise? a little I guess, you should work on those anima styles, chat
While interesting and seemingly easy to do, they are not there quite yet

Anyone tried the anima full finetune script yet? Or is the schizo reasoning on why loras are inherently flawed only applicable to sdxl?

>>8834459
that was always schizo and i never saw any real evidence (i.e. and x/y plot) that proved it, everything that was said pointed inexorably towards an armchair baker with zero actual experience

>prompt @koku on anima
>generic 2000s slop completely unrecognizable as him
>add game cg
>still slop
>add some of the more recent titles he worked on with a fairly consistent style that's more distinct than his earlier works
>STILL slop
What the fuck is this model's major malfunction? Fucking WAI has more of an idea what the guy's style looks like.

File: file.png (2.1 KB)

2.1 KB PNG

>>8834458
>>8834452
>mfw

>>8834484
it's aco

>>8834483
Yeah this happens to some artists for some arcane reason
Been waiting for some autistic anon to explain it for the peasants

>>8834475
Idk, from my own testing I always liked the finetunes more than the loras of the same dataset, but the effect kinda vanished when I tried to extract them.

anima Illuga lora doko

after you die from AIDS

>>8834490
it happens with the vast majority of the artists i care about for me
i'm (c/)hoping that it will be better with the full release

>>8834494
now this i can believe

Can any of you gents do a lora on One Piece
Like copying the pre-timeskip look or the wano artstyle
I love anima and can't go back to noob but I hate that it sucks for this

>>8834502
no

>>8834499
Like what artists anon, besides koku
I'll give some of them a try to ascertain it

>>8834502
kys shonenfag. how about you ask the power of friendship for a lora?

>>8834508
kek'd

>>8834507
why would i care about what you think? i know my use cases

>>8834511
Then stick your head up your ass you worm infected faggot

>>8834512
lol

>worm infected faggot
nah lil bro rly went and called blud a sakura stan

Why is chat so edgy today kek

sakura disrespect in the big 26 AINTNOWAY

>>8834502
it's kinda based that you only like the early part of OP but the style is fucking ugly and i refuse to make an ugly lora

>>8834514
What's with the worm thing? I only know of one /vg/ gachagame shitposter using it, and he's pretty new.

anima doesn't recognize my favorite aritst what do anons

>>8834507
Kobuichi and Muririn (they almost always draw together so it's one of the few cases where artist mixing works, just poorly), Shida Kazuhiro, Izumi Tsubasu, Amazuyu Tatsuki, Harada Takehito, Pochi (pochi-goya), Ishikei, Yaegashi Nan. All of these either look nothing like themselves or just terirbly bad.

>>8834483
just make a lora, unironically

>>8834520
>worm thing?
sakura having worms or faggots having worms?

>>8834523
No, it's way too bothersome.
>need to collect a dataset
>need to curate a dataset
>need to label a dataset
>need to use the right training settings (what settings are "right"? fuck knows!)
>need to wait however long it takes
>need to compare the different checkpoints to see which one is the best
Meanwhile if the arch wasn't stuck in 2023 and just had an edit mode, I could upload one singular image, say "do my prompt in this style" and be on my merry way.

>>8834520
It's a theory that homosexual and other degenerate sexual behaviors happen because of mind altering worms that use these behaviors as a way to propagate themselves
It's been shown in multiple studies that some worms will make people more promiscuous, much like the flu will make you more sociable before visible symptoms show up

im still using joycaption for captioning nat language, anything better come out recently?

>>8834530
is there a local joycaption? i'm still using eva03 large tagger or whatever the fuck it's called from 3 years ago

>>8834526
>Meanwhile if the arch wasn't stuck in 2023 and just had an edit mode, I could upload one singular image, say "do my prompt in this style" and be on my merry way.
there is not a single model that can actually accomplish this. No, NBP, grok and dall4 cannot actually do this, they're all too corposlopped. Even nai struggles to do this consistently because vibetransfer is garbage and so is their style reproduction/consistency.

1 year ago i mean

>>8834533
>zero-shot slop has zero shot at being good
next you'll tell me that I should have bought bitcoin in 2016

>>8834532
yeah its a comfy extension as well, it's fine I guess but pretty small model I wouldn't be surprised if something better was out now

Is there even a point of captioning images with nsfw nat language? with the te being censored and all? did the anima guy ever give us an example of how images were tagged?

>>8834537
So there are three ways to censor a model. You remove all loli, bestiality, etc from the dataset. Or you make it so everything captioned with these produces a blank screen or a "safety" image, like what Auraflow did unintentionally. Or the SAAS way, you run recognition models on the prompt text and on the output image, then block the user from seeing the output if it's tagged with "bad" stuff.

Third doesn't apply to local. Second I've never heard of anyone doing on purpose, it's unreliable and probably has negative side effects. And in the first case, the encoder still produces an output when encountering something it doesn't recognize, and you can align the Unet to whatever that output is.

He mentioned training with many different caption styles, including pure tags, pure nl, tags first, nl first and more. I don't remember the exact wording, check the HF readme.

>>8834522
I tested only a few of these in according to my own tastes
Harada Takehito worked great, Ishikei worked, although it looks more like his older stuff than his newer stuff, Pochi and Yaegashi Nan gave me generic slop

>>8834552
>worked, although it looks more like his older stuff than his newer stuff
That's a recurring theme too. You'd think "newest" would help but lol, lmao.

>>8834530
Beta one or alpha two? Is there a consensus on which one is better?
>>8834532
Autotagger is still goated for booru tags, and you can cycle through various models
But I really need one for natural language

>>8834557
link? or is that what it's called on github

taggui has a bunch of newer taggers as well

>>8834556
Yeah I've noticed that as well. I managed to get the old fate artstyle by using date tags and downright ABUSING the weight but I never managed to get the newer styles
>>8834558
Autotagger? Here: https://github.com/picobyte/stable-diffusion-webui-wd14-tagger
Be advised though, I couldn't get it to work with Forge Neo.

>>8834559
>go to the page
>look at screenshot
>UI isn't missing obvious features
>big ass preview pane
nice.

>>8834566
manual annotation is a bit cumbersome with it and i still prefer waifuset for that.

>>8834559
What's this taggui anon?

>>8834567
i'll see. i'm using boorudatasettagmanager, the russkies have made it a lot better in the last couple years

>>8834570
That's not pancho

>>8834571
https://github.com/jhc13/taggui

jordach status???

>>8834596
vanished in Bangkok after he gambled away the cascade donation money and then some

>>8834435
I've been keeping my noob mixes down to 3 artists (honestly, most of my mixes with 4+/5+ could have been shaved to 3 and barely affected the gen but cope) and been having success using 1 at full strength and scheduling the other 2 at 0.55
>>8834565
>Be advised though, I couldn't get it to work with Forge Neo.
Same. Crashed the entire thing for me so I personally switched over to >>8834574. I miss just having a tab to auto tag for me but I kinda like the ui on this program so I'll manage.

>>8834552
>>8834556
I've personally swapped from 'newest' to 'year 2025'. Another thing to note, experiment with samplers/schedulers. I've had styles look off with one that look almost artist perfect with another. Really depends on the artist too. Softer, more shading, not so flat works well on Euler A. The opposite does better with dpm++ 2m sde or er sde.

>>8834672
Very :>=ly

Why all DP (anal and vaginal) poses that show actual penetration are so weird

>>8834779
I'm finding that a lot of sex poses with 2 people are a lot lower quality than solo shots too. Any tricks I should be applying?

>>>/trash/81279915
:joy:

>>8834574
I might be retarded, what's the use case for this?

File: 00221-3541027986.jpg (982.7 KB)

982.7 KB JPG

>>8834786
Aside from inpainting, not really, I assume that you are already using all the tags you should be using

I want to test 4 tags and I want to see how they look in several different combinations, is there a way to do this without doing it manually?

>>8834794
>tags you should be using
Could you enlighten me? I'll cope for a bit you don't just mean "masterpiece, best quality, highly detailed"

>>8834790
I use it for gens where the image isn't from a booru, you know already tagged up, so maybe something I run into a trailer, twitter, or artist concepts of old games, etc. Run it there and have it do all the work for me and then just plug it into a ui to gen with.
I'm sure most use it for compiling datasets for loras.

>>8834786
Are you talking anima or illu/noob? The latter clearly is a vae limitation. You're stretching that shit thin the more you pile into the image. Hence why when you have /hgg/'s bread n butter, 1girl solo looking at viewer close-up, it'll effortlessly look crisp. As for anima, it does a lot better due to the larger vae but you'll still have an issue here and there which you can chalk up to the preview model and the fact it's only been trained on 512x512. Or at least that's the going narrative.

how do you find inspiration for poses and then figuring out how to type it out in a way that anima understands?
I'm struggling with coming up with creative stuff so I thought I'd try looking at some references but I have no idea how I'd even begin to explain how these pose are set up
I love natural language but I think it's harder to gauge what works and doesn't, unlike tags

>>8834843
Oh that's a shame. Do you think a custom VAE would help?

>>8834845
>I love natural language but I think it's harder to gauge what works and doesn't, unlike tags
A little yeah. With noob, I'd throw in a shit ton of tags that I think would accomplish the image I saw in my mind's eye and when it worked, great. When it didn't, I'd mald and keep at it until I got something close or something completely unique.
Anima, on the other hand, you can do the same but the nl support gives you even more control so any failings are truly a skill issue or lazyness. Describe to it like the lobotomized model it literally is what you want in one sentence, throwing in actual tags in the process, and you should get what you want mo sweat. I think camera angles is the most difficult thing about it since that's always influenced by the resolution.
>>8834847
Nope. Those normally make it worse, especially with inpaint. The vae is only 4-channel so there's no getting around that limitation.

>>8834800
x/y plot in scripts section of a1111 based uis (at the bottom)

>>8834832
https://danbooru.donmai.us/wiki_pages/tag_groups

>local music gen is now a thing
>and it's actually surprisingly really lightweight
>but the trainers for it are garbage
>also captioning is absolute cancer
just kill me

>>8834857
NAI status?

>>8834859
absolutely fucking dead in the water. so dead they're not even a zombie or a skellington, they're just dust.

i hate nai so much

Based

i will never have 25 dollars...

>>8834857
please elaborate
i want to musegen for memes

>>8834857
Is voice cloning, particularly music lyrics replacement, still where it was in 2023? I looked it up a bit and while normal voice cloning/TTS did make some progress with stuff like Qwen3-TTS, I couldn't find anything beyond RVC for lyrics replacement.

>>8834875
there's a thread on /g/ >>>/g/108164777
the model came out a few days ago. The UI and training portion of all of this is basically vibecoded trash.
But it's in general really lightweight. A 12gb 3060 can generate two 90 second tracks in like 35 seconds(read: you can generate faster than you can listen to it on a $200 (used) GPU). It can do lyrics if you want, but captions are all LLM generated wordy garbage, lyric sheets require really specific syntax and it's all NL garbage.
The trainers are bad. Really fucking bad. It recognizes that there's a u-net and a TE(and the ability to effectively train unet only) but it only gives you the option to variable a single, nondescript LR that I have no idea if it's actually for said u-net/TE or if it's something retarded like the GLOBAL LR that gets set for the optimizer.
>>8834877
honestly haven't played with it that much in that respect. Right now I'm just dipping in for the experience and to learn how it all functions. Someone posted something like a lyric replacement in the /g/ thread but it still sounded pretty TTS robotic tier.

>>8834881
>cumrag AGAIN
*yawn*

>>8834881
>only one LR setting
end me

p-pancho aniki... *cough* ,, t-that's 2 updates we n-need.. o.. onegai.. *cough*

>>8834881
>Generate a full 4-minute song in ~1 second on RTX 5090, or under 10 seconds on RTX 3090 with ComfyUI
Why can't vidgen be like this?

>>8834881
the samples sound like early Suno
which i guess is pretty good for local

>>8834882
there's a gradio UI, you can completely ignore cumsoft. The better trainer is commandline only.
>>8834883
it gets worse.
here's a very brief list of my complaints >>>/g/108194641
oh also it has two different architectures for training. bogstandard LoRa and lyco-type! But it's. Uh. um. It's LoKr.
>>8834884
just be like me and completely ignore anima until the full res model is out. Maybe by then it'll have significantly improved training support, too.

>>8834887
lmao okay i guess i'll just sign up for suno or something

i already like anima more than sdxl desu
except for no upscaling

just upscale with nai

>>8834891
sorry chuddite my only $25/mo sub is a t3 to my oshi

>>8834889
I'm enjoying the vibrancy and detail from a 1.3-1.4x hires anima gen compared to the 2-3x cn upscaling coping of sdxl thoubeit

>>8834893
what kind of retarded psychopath 2-3x upscaled instead of 1.5x or something, you need help

>it's unheard of to upscale a shitty 1024x1024 sdxl gen to 2048x2048 and higher
Can't you civjeets just accumulate buzz to upscale or some shit if you can't afford a gpu?

>>8834896
i have a 4090, you are just stupid

>>8834897
silence gpulet

No you don't, Rajeesh

File: 1771592838.jpg (1.3 MB)

1.3 MB JPG

10 minutes on a gtx 970, being poor is no excuse

>>8834485
box please?

also, what are good negatives on anima

>>8834945
i did some testing a while ago and imo it should be pos/neg

masterpiece,
(ye-pop deviantart:1.2), worst quality,

personally i also add (@hero neisan:0.2) to negs because it's intensely sepia and fat two things i do not want

>>8834526
This couldn't be any more accurate

>>8834948
>(ye-pop deviantart:1.2), worst quality,
should there be a comma

>>8834952
i won't say it doesn't matter but i'll say it's hard to predict how commas affect style negatives

>>8834953
russell noted spaces matter a lot so i would assume it is important. problem is the example caption in the model card has them on first line with no comma

>>8834956
good to know. i found out the other day that the TE is sensitive to misspellings, i was tying in a character's name wrong and it didn't work at all which would be a problem at all with a CLIP TE

wouldn't*

>>8834952
>>8834953
>>8834956
>>8834960
Comma:
https://files.catbox.moe/gl8vjb.png
No comma:
https://files.catbox.moe/evk53s.png

>>8834962
images are broken for me

>>8834962
catbox is fucked, use litter
and anyways i'm gonna test it myself

File: no comma.png (772.7 KB)

772.7 KB PNG

>>8834965
Fucking catbox.

>testing on a single fucking image
begone retard

File: no comma.png (798.5 KB)

798.5 KB PNG

>>8834967
I'm retarded, ignore the filename in the previous post, it IS with a comma. This is without.

>>8834936
>gtx 970
Had that ages ago. Still remember the entire memetroversy behind the 4gb but not really. Served me well enough until the 20XX class.
>>8834948
>and fat
We don't want you in the ojisan club.
>>8834967
>>8834970
>the comma is the catalyst to a sweater
What did he mean by this?

>>8834910
There is no way for that poor girl to pretend to be Taihoe on any way

>>8834982
That's the joke.

>>8834983
Stop bullying womanlets

>>8834986
why he mad

>>8834989
You ever finna nut after a coke and alcohol binge, anon? Sometimes you just wanna finish but the rest of your body does not.

>>8834989
he's focused

>>8834991
sorry i'm not european i don't do coke

any hopes of forge couple getting anima support?

>>8835007
Why would it need special support?

I'm finding it pretty hard to control breast size on Anima
I would love to see an example of how body shapes are tagged in NL because I feel like with NL it should be easier to control the shape and size of breasts
If they've even tagged breast sizes, because I seriously hope it's not just tagged "breasts"

>>8835029
If you feel the need to use tags other than large, huge and sagging, you have lost control of your life

>>8835030
medium is premium you dog

>bake anima lora (2 hours)
>not working all because of scheduler in constant rather than consine.

Anyone got a good formula by now? it takes so long where I can literally bake an illus/noob lora in less than 30 mins.

>>8835032
>it takes so long
maybe try the finetuning script. finetunes were a lot quicker than loras on sdxl if you had the vram

>>8835030
this >>8835031
but it's also about shape, I want to be able to define the shape of the breasts but it's just giving me different sizes of sagging breasts despite negging sagging breasts

>>8835033
it's anima I need help with not sdxl....

also I only have 16gb vram, anima takes only 8 gb with gradient ticked on. no idea what a finetuning script is though.

File: 1771604395.jpg (441.3 KB)

441.3 KB JPG

>>8834974
old and busted but it better last me another decade because i'm not buying a new gpu any time soon
it can still spit out something small in like 3 minutes, though

>>8835007
You don't need it. You can just say "the girl on the left / the girl on the right" or any other image composition thing you want. Camera angles are kind of annoying but relative direction works completely fine.

>>8835029
It works the same as normal. flat chest > small breasts > medium breasts > large breasts > huge breasts > gigantic breasts > hyper breasts

And you can use all the modifiers like sagging, bouncing, hanging, breasts apart, breasts together. It's also based on artist, obviously, and just use negs if you hit some weird place in the latent space.

skill issue

>>8835032
why would "good" mean "30 minutes?"
i never baked anything good that quick on sdxl

>>8835056
It's not the same. Nl still bleeds and quickly reaches its limit once you want to separate a lot of concepts

>>8835076
well i have noticed some stuff to that effect but i put it off to 512x res

sage and reported

who

>Nearphotison has 15 posts on danbooru
>Still works flawlessly
Where does it come from

>>8835106
as long as it's clear enough artist loras can work with that few images so why wouldn't normal model training also work

>>8835106
danbooru has dumb rules
951 pictures on gelbooru

also
>look it up
looks like dogshit

>>8835106
it looks awful though

let's also appreciate how fucking awful it is that it's cringe ass text over cringe ass shounenshit characters.

i won't kink shame you but i WILL taste shame you

>>8835106
so, this is the fabled western anime

>>8835110
LoRAs have their entire file size budget dedicated wholly to just one style/concept; checkpoints have to somehow fit the ENTIERTY of Danbooru inside like 4 gigs, which usually means that any tag without at least a good several dozen examples will simply not be learned. And something tells me that dividing 4 gigs by the number of unique tags on Danbooru will land you at a FAR smaller size than a LoRA's typical file size of anywhere between a few dozen and several hundred megs.

>>8835118
yeah i was bein retarded

>>8835106
https://www.deviantart.com/nearphotison

>>8835121
oh
oh no

>>8835118
4 182 218 360 bytes (size of the Anima checkpoint) / 313 235 (number of lines in the 2025/09/19 Danbooru .csv) lands you at just 13.35 kilobytes per tag on average. That's the ballpark of embedding (remember those?) file sizes, not LoRAs. If we make the fairly safe assumption that the vast majority of the tags are below the knowledge threshold, meaning that a small fraction of the top tags takes the majority of the file size, that leaves even less space, and much less at that, than 13.35 kilobytes for any one of the far more numerous smaller tags.

your prompts need to be blessed by a japanese person
to not be western prompts
the prompt priest

>>8835127
i saw their prompts in the 1.5 days
these niggas are a superstitious clownshow prayin to the latent space like it's the god of the fuckin mountain

>>8835127
clap clap clap

start putting a little japanese flag on your gen to make it japanese :)

File: {60B48FEA-3A0C-41A1-B3F5-876A195880B0}.png (45.2 KB)

45.2 KB PNG

I think my finetune config might be chopped

I am trying to prompt two views of the same scene as I want a view from the front to also see the girl's face during sex from behind but it's messing up whether it's turning the girl's head 180 degrees on the back view or completely failing at giving me a front view at all
>there are two views defined by an even split in the middle of the image. the view on the right is a pov from the man's perspective as he grab's the girl's ass with his hands sinking deep into the skin of her ass cheeks while fucking her. the view on the left is the same scene but from the front of the girl featuring her expression and reaction to the sex with the man standing behind her as he is fucking her pussy from behind with his hands on her ass, sex from behind, the man's head is head out of frame on the left view, the setting is the same in both instances in a dimly lit hotel room and the girl is leaning over a desk with a desk lamp on it her breasts resting on the desk
This is my shitty attempt, maybe someone else can come up with something better?

unc is typing out his life story and thinks 0.6b can handle it :skull:

let's see your prose then, big guy

>>8835121
Jesus I haven't been to DA in like 15+ years and got jumpscared by that layout.

>>8835145
Instead of trying to fit everything on the same prompt you could just overlap two different gens together

>>8835153
I mean yeah, I could just do two gens but isn't the fun with NL that you experiment with how much you can prompt for?

A year later and here we are just writing...

>>8835145
The approach I've been having is to use first an NL description of general aspects of the composition and then do a separate line for booru tags to support it and emphasize individual features (even in a multiple views scenario) and changing one or the other up depending on how the gens do/don't reflect what I'm aiming for.
So in this case I would suggest rephrasing most of it into 1-2 sentences and all the small stuff relegated to tags as you normally would do.

>>8835159
Yeah I'm currently rewriting the whole thing as it got a bit messy

>>8835154
That's a good point yeah, perhaps try to do a list of things you want on the gen just for the lulz

>>8835149
i should actually make at least one of the doujin ideas i've had floating around in my head now

>>8835110
I don't think that's the case
>>8835111
So, do they use gelbooru as a database?
>>8835112
>>8835113
kek it is pretty terrible, but it helps with the anatomy if you schedule it out at 0.2
>>8835121
Let us hope THAT didn't make it into the database lmfao

I condensed it to
>sex from behind, vaginal, grabbing another's ass, deep skin,
>the man is fucking the girl in the ass penetrating her pussy with his girthy penis while he grabs her ass with his hands sinking deep into her soft ass cheeks,
>the image is evenly split in two views of the same scene by a thin black line going vertically through the middle with the view on the right being the pov of the man and the view on the left is from the front
It just wont work. With this prompt it's giving me the pov sex from behind on the right but the left side is just the girl standing, facing the viewer
Why would the danbooru tags not apply universally

>le multiple views tag

>>8835185
you need to reread the wiki my man

adding
>multiple views
does nothing

>>8835187
>the wiki
What wiki??

>>8835193
https://danbooru.donmai.us/wiki_pages/tag_groups

take a look at image comp and camera angle tags and see how they are associated with actual images, then apply that to your prompt

Just list the tags lil bro

>>8835176
We know DA was a part of the anima dataset, we know he has 700 pics on DA and 15 on Danbooru. Safe to say it's included.

>>8835194
I am obviously aware of the danbooru tags. The ones that would help me would be multiple views and maybe 2koma/comic but neither do
So unless someone else can successfully prompt for what I'm trying to do I think I'll just have to write it off as not possible

>>8835198
You're missing a lot of tags for what you want to achieve though
For instance, doggystyle and missionary

>>8835198
skill issue

>>8835201
Half the gen works though, sex from behind is doing its job as the right side is giving me a pov of the girl getting her ass fucked. Meanwhile the left consistently features no sex at all happening

>>8835202
Post yours :^)

so ive been out of the loop on anima, you cant mix styles? why? isnt that like the whole point of image gen goonery?

File: 00022-4170550367.png (965.8 KB)

965.8 KB PNG

>>8835198
nta but i felt like trying it, something like this?

>>8835222
why is it so blurry?

>>8835218
it's just annoying, not impossible

File: 02199-823266104-yuming li, (traditional media, oil painting _(medium_), watercolor _(medium_), impasto_1.2), (ringodrawing_0.8),_best quality, a.png (2.2 MB)

2.2 MB PNG

Been out of the loop for a little under 2 years now. What's the best way to prompt two different characters with their own lora? Last I remember the best way was to use regional prompts, is that still the way to go or is there a better way now?

>>8835241
most new models have pretty good characters out of the box. the actual best solution is to train a lora with the pair

File: 00016-2726385220.jpg (958 KB)

958 KB JPG

>>8835241
it depends on which 2 characters and what type of interaction you want, it can be as easy a just pure prompt or using several hacks like controlnet and inpaint to hell and back

>>8835245
he did say lora characters

>>8835241
if you're not training them yourself it depends on how they were tagged, they might leak features onto each other
there's also regional lora scheduling, a little harder to work with and slows down gen time

>>8835224
It's NAIma

>>8834967
>>8834970
doesnt seem too different

File: 1753495008725952.png (1.1 MB)

1.1 MB PNG

>>8835185
you're still over-complicating it. remember that it's a 0.6b model and try something like this instead:
instant loss, 2koma, multiple views,
top panel:\(rosemi lovelock is looking smugly at the viewer, 1girl, solo, rosemi lovelock, nijisanji en, grin, looking at viewer, hand to mouth, mesugaki, outdoors, portrait\)
bottom panel:\(1girl, 1boy, mating press, plap, sound effects, boy on top, fat male, deep penetration, vaginal, anus, uncensored, leg lock, impregnation, cum in pussy, hetero, folded, on bed, indoors, bedroom, full body\)

>>8835308
>that text and sound effects
your 16ch vae model, local saars

>>8835308
This is exactly how I tag 2koma and have no prob.
>quality, artists, # of players, 2koma, general tags not easily conveyed in nl
>top panel is yada yada yada
>bottom panel is yada yada yada
It's a lobotomized model so speak to it like a fellow indian anon when describing things. Don't be too verbose.

>got a new GPU
>this completely broke reforge because of course it did because lmao cuda problems
>had a reforged neo install I didn't bother with the other day because it was requiring cuda 13.0 and I wasn't updating to that and breaking anything, but since I have to use it now anyway whatever
>get it installed and running for all of about 10 minutes until I start trying to install extensions
>restart UI, EVERYTHING COMPLETELY FUCKING BREAKS
>can't figure out how to revert it, deleting extensions does nothing
>try and just do a clean pull and restart
>now it's throwing an error requiring python 3.13
I want to die

>>8835322
kekerina of time. I ran into the same shit trying to initially get on Neo. Just a straight up git clone and install with requirements was throwing up errors related to python 3.13 as well. I'm sure I had a separate install of it which I verified so I changed the env paths to reflect it and then it started throwing shit about numpy, some audio engine (I imagine due to the wan support) and some other minor shit. Raged and deleted everything. Do a bit more research and see some shit about Stability Matrix basically doing everything for you and the option to swap ui's on the fly. Cool. Try it. Some python error again. Fuck you. Uninstall everything go to the Neo page and actually read this time. Choose the uv install route. Everything turned out better than expected.
Try the uv route, anon.

>>8835326
thanks, that actually worked.
also, neat, gens take about half as long as they did before. wasn't quite the goal/purpose of the GPU upgrade but still nice I guess.
maybe now I'll actually fuck around with anima tomorrow.

>can't pay for Anlas via Paypal anymore
It's over....

OINK OINK

ixybros.. our genuine response?

bros... any advice for anima gens?

>>8835357
Just... write

>try out anima
>try to upscale to see if it actually is fucked or if it was just people being retarded and/or a fucked cumrag upscale implementation
>upscaling works perfectly fine
wow it really just needs controlnet tile and for actual trainers that matter to implement code for it in addition to it actually being trained the full way through since this feels pretty rough at the moment.

also I'm guessing upscaling issues people were complaining about is just another failed cumrag implementation(who could have possibly seen that coming), since 1.5~x upscaling with a decent amount of denoise works perfectly fine, it just has the issue of not having a controlnet option to reinforce the composition.

you are you all obsessed with upscaling

because i am an ARTIST

>>8835374
it's a problem with portrait resolution
https://github.com/Haoming02/sd-webui-forge-classic/issues/763

>>8835222
Oh yeah, very nice! How did you prompt for that?

>>8835308
I'll give your style of prompting a try, but the above anon already solved it. Hopefully he can share with us how he did it

>>8835381
I am quite literally upscaling images that are taller than they are wide.
it's a bit rough around the edges and it has some really obvious resolution driven anatomy issues but it's functional and isn't spazzing out.
https://files.catbox.moe/6ubki7.png

>>8835392
what happened to her groin :(

>>8835374
>>8835392
Anon, plenty of us have been upscaling by 1.3-1.5x. I personally haven't tried beyond that myself but the issue seems to be 2x and beyond. Another anon is coping meanwhile with ultimate sd upscale.

>>8835395
cope anon also specified the issue starts at a little over 2 megapixel

>>8835357
elaborate

>>8835410
quality tags/negs.
prompt structure. sampler/scheduler/cfg

https://huggingface.co/nightknocker/cosmos-bert
>BERT for Anima/Cosmos.
>This is not an adapter model, but rather an early replacement for the T5/Qwen model.
>This means that the T5, Qwen, and LLM adapter files are about to say goodbye.

>>8835412
what is this? custom text encoder for anima?

>>8835412
lil bro forgot about ernie

>>8835415
never beating the allegations

>>8835411
see >>8834948 and replies (i do think no comma is better)

prompt structure is the same as before except you can use NL. so think about camera angle, composition, and body shape first, then do whatever.

note that the advice on the anima page https://huggingface.co/circlestone-labs/Anima about what order to put stuff before the main prompt is mandatory, shit will break if you don't put things in order. also note that unlike CLIP typos can cause tags to just be ignored

>>8835412
so can i just put this in the TE box or what? also
>file named model.safetensors
my retard alert is blinking

>>8835418
>file named model.safetensors
that's just default diffusers naming?

>>8835420
yes exactly. it means they didn't even take the 5 seconds it takes to rename their file to what it is.

>>8835383
sure thing
https://files.catbox.moe/zd8fmr.png

you know, highresfix works not so bad on anima

>>8835423
awesome

>>8835423
now this is some good prompting

>highres fix upscale 2x, 0.3 denoise, no upscaler "none" option
>send to img2img
>denoise 0.32, multidiffusion tilesize at base resolution, mix of diffusers option, overlap of 512 for sdxl, rescale 2x, random seed, click orange button
>open result in krita, downscale 4x image, erase any mistakes using lower resolution image, save new 2x resolution image that looks perfect with no double bellybuttons or weird shit happening in it with more detail that took 2 minutes more to make
>post
highres fix works great in a1111 too have been using it forever, i dont get why anyone would use an upscaler alone when they all just take longer for the same or worse result as "none" in sdxl

File: ComfyUI_temp_qquqa_00015_.jpg (178.9 KB)

178.9 KB JPG

Thanks for all the help, bros
Anima truly is wondrous

>>8835431
what the fuck are all these extra steps lmao

https://files.catbox.moe/sgd8nb.png

>>8835395
only a retard would bother trying to upscale that much though. it doesn't matter what you do it's always going to look like trash unless you either look at it on a phone or you have it zoomed out at half scale. Maybe if the models were trained at 2048px it would be functionally fine but we've seen how that sort of shit goes with illustrious v2.

>>8835431
>>8835433
"none" upscaler is Lanczos
Multidiffusion is tiled upscaling

>>8835428
It's simple prompting as it should be. I keep saying some of ya niggas getting too shakespearean with this lobotomized text encoder.

is the default save precision of kohya fp32 or why are my finetune safetensors 8gb?

>>8835437
i am aware

>>8835436
>only a retard would bother trying to upscale that much though
It's easier to upscale to absurdres and then downscale it back compared to inpainting little details that get destroyed by 4ch vae. Honestly anima didn't feel like much of an improvement when it comes to intricate outfit details.

>>8835451
>Honestly anima didn't feel like much of an improvement when it comes to intricate outfit details.
well the positive effect of the 16ch vae while genning is destroyed by the shit upscaling and the positive effect while training was destroyed by the 512x dataset.

>>8835451
i disagree, besides the obvious degradation that training at 512x made, i've done a bit of testing on highly specific concept loras, and the ability to pick up details is greatly improved, better than sdxl after months of testing and dataset curation. now some of this could just be simply that anima's dataset is superior but i don't think so.

>>8835451
what details. to upscale to resolutions that large you have to nerf denoise to miniscule levels and all you get is nearly exactly what you started with at low resolution but fuzzier and with a lot of halo/glowing between linework. You're not introducing or improving any details over what you had at the start.

>>8835454
>nerf denoise
tile controlnet

>>8835455
does not yet exist for anima
have we strayed into the theoretical zone

>>8835455
if you're setting controlnet tile to .9+ weight and 1 end step you're not really doing much with it.

>>8835457
personally i like the weight fairly high but drop it off at like .6 or somethin

How do you guys manage 1.5x upscales on anima? It just artifacts for me no matter the settings

why do you want to upscale? what are you going to do with the image?

>>8835484
>what are you going to do with the image?
fap

>>8835485
why do you need to ultra zoom in on your gen to fap

>>8835486
why wouldn't I?

>>8835483
Just read the thread

You upscalefags just wouldn't get the Dell E773c 17" Monitor anon.

Is being anti-upscaling the new contrarian brainlet take after anti-lora, anti-controlnet and anti-regional prompter?

>>8835501
I think so, everything for the sake of those sweet (You)s I guess

>>8835501
Upscaling isn't free - it costs time. On SDXL that wasn't much of a problem since it's so fast AND upscaling was a massive improvement over no upscaling because of the compressed VAE, but on Anima that's both slower and much better at details out of the box, its relatively smaller benefits are much harder to justify when juxtaposed against its much harsher downsides. You're basically paying more to get less benefit.

In order for the model to give good outputs right off the bat without upscaling/detailing, it has to be a lot bigger. Z also needs upscaling while Qwen Image can give nice results right away

Anima actually murders details because of the 512x training. You can see this pretty easily by prompting any character with clothes writing and comparing with what NAI 4.5 would give you.

>>8835509
That's only temporary, it's ought to get much better once it's trained on 1024p. Still, even the way that it is, it simply artifacts less and produces much more coherent hands when they're far away from the camera.

>>8835512
That's the bare minimum one would expect from a 16ch vae model, but sure.

>>8835512
>once it's trained on 1024p
ETA never ever

>>8835512
Were there any updates on this after the preview release? Any info on the progress?

>>8835517
the only thing he's said is that it'll be trained for a long time at 1024px

>>8835518
i will wait warmly

https://files.catbox.moe/2d4xkm.png

anima is kind of neat in dumb ways, I suppose.

>>8835543
kino

drawfag lora when

>>8835604
he just like me frfr

>>8835604
Why is a Ukrainian pulling Burnice's panties?

>>8835604
Based Napoleon enjoyer.

sirs?

>>8835607
hehe

>>8835625
Yes?

>>8835638
good morning

>>8835641
it's 5 fucking pm and I am dying of heat wtf is wrong with you chat

>>8835642
do not redeem the upscale

help

>>8835666
what with

>>8835667
not upscaling

any new fast anima lora presets working? tired of baking overnight only to see it never works well.

>>8835667
gooning

>>8835670
actually on that note my first two bakes worked fine it's just in an "meh, ok" state, which took ~2hours per full complete bake at 3500 steps, but will changing from 1024x1024 to 512x512 help with the quality or does nothing?

>>8835672
i only got worse results when i tried 512x512 but maybe that's just me. it's also much faster so you might as well try it out if you're curious
i've also only gotten results at the "meh, ok" level so far, although there are definitely things it does much better than sdxl models

>>8835670
i just copy pasted my settings and it went from ~1:50 to ~2:30, results mostly good

not gonna spend a ton of time baking since it's just a preview model, i still have a massive backlog of datasets to prep

>>8835671
elaborate

>>8835677
hentai

>>8835678
like what

>>8835679
girl

>I can taste... The essence of your suffering - it is sweeeeeet.

File: 20260221205651-1619422409.png (1.2 MB)

1.2 MB PNG

>>8835680

>>8835692
thanks

>>8835692
>nipple piercing
based

>>8835692
You never had the makings of a oji-club member.

File: 20260221212420-2426971709.png (1.1 MB)

1.1 MB PNG

>>8835699
i respect women irl
it's only in the world of imagination that i rape and enslave them

>>8835702
Is that one of the roach men from Terra Formars?

>>8835702
>i respect women irl
shiggy diggy

>>8835736
pretty much yeah

can you make the male younger

>>8835828
chat, is lil bro the new meta?

>>8835828
Cringe.

can you make the female younger

>>8835819
cute

>>8835831
her age?

>>8835834
25

>>8835835
too old

>>8835836
I liked your performance in titanic

>>8835836
unc, current meta is that pubescence ends at 32