Premium Service EditionPrevious Thread: >>>>8834450>LOCAL UIforge neo (anima support): https://github.com/Haoming02/sd-webui-forge-classic/tree/neoreForge: https://github.com/Panchovix/stable-diffusion-webui-reForge | Stable commit: 6964cebComfy: https://github.com/comfyanonymous/ComfyUI | https://comfyanonymous.github.io/ComfyUI_examples>NOOBAI-XLhttps://civitai.com/models/833294/noobai-xl-nai-xlhttps://huggingface.co/Laxhar>NOOBAI SHITMIXES102d custom: https://civitai.com/models/1201815?modelVersionId=1491533291h: https://civitai.com/models/1301670/291h>RESOURCES4chanX Catbox/NAI prompt userscript: https://rentry.org/hdgcbEasy booru tag export userscript: https://github.com/Takenoko3333/Danbooru-Tags-Sort-ExporterTags: https://danbooru.donmai.us/wiki_pages/tag_groups | https://danbooru.donmai.us/related_tag | https://tagexplorer.github.io/#/Inpaint: https://files.catbox.moe/fbzsxb.jpg | https://huggingface.co/Wenaka/NoobAI_XL_Inpainting_ControlNet_FullUpscalers: https://openmodeldb.infoBooru: https://aibooru.onlineControlNet: https://rentry.org/dummycontrolnet (OLD) | https://civitai.com/models/136070 (look at the green links in the description)>TRAININGGuide: https://rentry.org/yahbgtr (WIP)Anon's scripts: https://mega.nz/folder/VxYFhAYb#FQZn8iz_SxWV3x1BBaJGbwTrainers: https://github.com/derrian-distro/LoRA_Easy_Training_Scripts | https://github.com/bmaltais/kohya_ss | https://github.com/Nerogar/OneTrainer>NEWShttps://huggingface.co/circlestone-labs/Animahttps://huggingface.co/ChenkinRF/ChenkinNoob-XL-v0.2-Rectified-FlowOP Template/Logo: https://rentry.org/hggop/edit | https://files.catbox.moe/om5a99.png
do not let /hdg/ troglodytes infiltrate
Too late, local lost
...the losing competition
uhhh anima gens?
>>8839919Anilingus is disgusting and gay.
no
>>8839935low test insecure beta
>>8839935WEIRD comment
did they post the arena thing's models? feels like an ad for the top 5 other than anima.
>>8839947wonder if he's ever going to post them.and TO BE FAIR, the only reason anima is that low on cond leaderboard is because prompts there are just booru tags, undermining the real point of anima (llm text encoder).
>>8839935>it's gay for a girl to lick a guy
>>8839954so true! it's also completely straight for a guy to crossdress and get pegged by his girlfriend :3
it's so fucking gay when a girl completely subjects herself to the man's pleasure and is even ready to degrade herself licking his asshole. it's basically same as getting raped in the ass
Hard pill to swallow, but if you hate women enough to want to see them degraded then you're gayer than people who have gay sex and you probably have manboobs
>>8839958what about the maso subs
>>8839964The female equivalent of a cuckoldThe woman who cucks her husband still hurts him even if he says he likes itLikewise with the opposite, though it's less egregious if you keep it light
Only missionary position sex with the sole purpose of recreation allowed here.
>>8839956easy on the projection there
Fellas what's more gay, fuck a cute boy or let a girl lick your ass?
>>8839931
>>8839970Based as fuck>>8839974Objectively the latter
gay status?
>>8839978chat doesn't like girls for some reason, which explains the lack of 1girl gens here
I wish you faggots were as gay as you pretend to beWhere are the 1boy gens
they are for my eyes only
for me it's 1frot, small penis, foreskin
Ayakon spotted in the wild>>>/vg/559094430
so, these are the fabled quality /hgg/ discussions
>>8839987>still on pony in the big 26i kneel
>>8839988chat is just being a little silly as usual
>>8839992can you gen her rimming an obese guy?
>>8839994that would be gay so, no
BASED
>>8839995retard
huehuehuehue
lose weightdrink more water
i thought mexicans only have coke
>>8839992can you gen her rimming a horse or churl?
Shit I wanna see a churl on all fours
>>8839995Truth-teller
>>8840004rimmed by all the genshin sluts simultaneously
skin the color of excrement? yes please!
What about frost trolls?
>>8840006inviting (you) to rim him
Why is chat like this
Maybe I should gen it
Ok, I have a question that is hard as fuck to solvesome character loras only work when the trigger word is in prompt, like pic exampleif I'm downloading a lora, what should I seek (in descriptions)?but if I'm training a lora, what should I do (in configs)?I have lots and lots of characters loras, and some have strong triggers, others do not, the ones that have strong triggers always work betterpls frens help me with thisbtw any good guide in how to train a lora will help a lot, I've never found a good guide
>>8840020already answered in the previous thread
>>8840023Sorry just saw that, ty for the helpI will try to add the trigger word firstalso still needing a new and good guide to bake loras, all I found are indians or chuds trying to bake real life porn loras, and that sucks
we need hgg2...
>>8840028just use tripcodes lol
that play is a little too rough nglhttps://litter.catbox.moe/dt1g1s4oqdp90wur.png
Good morning, fellow sirs :)
we are so back, gang
that is not churl
>>8840039this, not black enough
newfag here, what's the difference between /hgg/ and /hdg/?
>>8840048why bother making IA if you cant tell the diference?oh you must be a NAI user...
>>8840049Hnnngh~ amazing lamia! Love her!
>>8840049his age?
belly bulge is is for retards
What's with all the low test beta comments here lately
>>8840065Gorgeous stretched tum-tum~
>>8840072/hdg/ died so the troglodytes moved here. we need another split
we need a thread for pedos, fags and indians, so the white man can post onee-shota here in peace
Is NovelAI good? I'v seen a lot of people using it.
>>8840080it's the model of choice in /hdg/, you should ask there
>>8840080quite terrible in need of an update, only thing it has going is cloud access inpainting
The government doesn't want you to know this but you can legally rape every Hilichurl you see, they have no rightshttps://litter.catbox.moe/zi7y58po058c92fi.png
churls do the raping
>>8840091What a mighty peculiar churl.
>>8840091based
hmmm, i'm gonna gen otoko no ko churls
https://litter.catbox.moe/aoa5b3mmihnqan6k.png>>8840097It was a bit of a pain to gen the ears and horns as separate things but it gave me an idea>>8840099Noooo you can't do that anon
How do I achieve something exactly like this?btw thats a churl cock if anyone ask
kek
The reason churls use masks is because they don't do oral, they are exclusively anal only
>>8840108based cuckbox rejector
>>8840109I only used it because of rule 3 lol
>>8840111anon...
https://x.com/GA006_/status/2028309538223669401How are his gens so clean bros?
wtf
>>8840123I don't know how the japanese do it, there's a solid number of them that make magic with local models
When will novelai make its own vidgen?
>>8840125>japanese>Pyccкий
>>8840123pick a style that isn't 2.5d / gradient shaded so we assume it's artsy looking
>>8840049kino....
https://litter.catbox.moe/3w2rulb3myi7zx5a.png
>>8840123it's nai
>>8840127they dont care about video gen
>Claude>Doesn't care about image gen>Best text model in the world>NAI>Doesn't care about vidgen>Best image model in the worldmakes sense
>>8840154>Best image model in the worldNAI architecturally is pissy shit compared to models like Seedream 5.0.
>headcanon>headcanonmakes sense
What would mein Führer do on this situation
where is the anon with the baked nangong for ilu?
where is the cunny
When hgg2 goes up finally, how long before the golden age ends that time? I'm thinking #4.
>>8840211You act as if this general had a golden age. Its peak was #1.
hgg2 waiting room
>>8840103You have turn bibicipred on ;^)
>>8840162Hit me with a hitler lora for anima and I'll tell you
Why doesn't 'jinbaori' work on anima, it worked on noob
why are /hdg/ brownoid FAGGOTS here?
>>8840255I do not see any
>>8840256probably means you are one
>>8840257:^(
anima sampler/scheduler/cfg?
>>8840294Euler A normal 4
>>8840316Don't forget score_9 in your prompt btw
SHOW ME THE CUNNY
>>8840330in the negatives of course, forgot to mention that
first for novel ai
russel posted on ldgapparently anima preview2 "soonish"
>>8840353>preview2omg bruh :wilted rose: :broken heart:
:skull: im dead
aintnowaybruhfuck my chungus life
ywnbaz, unc
aibros, are we safe?https://x.com/JIMMYEDGAR/status/2028606326205567375
>>8840373they're talking about nai v5
>>8840374>analtan:skull:
post black umamusumes getting bleached
umas are for horse cock>https://files.catbox.moe/e0xcqo.png
>>8840391can you make the stud darker
Good morning sirs :p
>>8840397wrong thread, faggot
>>8840397right thread, straight male
>>8840397thank you, she's cute!
what the FUCK
her age?
young
her sloppiness levels?
out of ten!
>>8840451>>8840449>>8840448>>8840447>>8840446Go back to /hdg/!
this, churls doko
>>8840452you missed the scat fetishist
>>8840456it's /hgg/ culture, tho
>>8840457/hdg/ actually
>>8840458yeah, as shitposts/hgg/ always had unironic okinawan males, churls, bestiality, half meter dicks and other faggotry
>>8840460which are all fine because it's not poo poo fetish shit
>>8840461it's literally just as low
>>8840461sure, anonsame cope as changing one letter in the general name
>>8840462no. because poo poo shit is also fetishized irl
>>8840464you say that as if bestiality doesn't have the same audience
>>8840466as far as i'm aware not really
>>8840468explain churls then
the difference is that you need to pretend churls are people
>>8840469so you're saying all those women reading beast goonslop is the same as them wanting to be bl****ed?
>>8840470this, a black male isn't a black male because of 4 fingers and a tribal mask
https://litter.catbox.moe/pb1qjv.jpeg
>>8840470don't you also need to pretend blacks are people too?
>>8840474true
>>8840473trvke
https://litter.catbox.moe/9b7cm2.png
Just getting back into genning and I see Neo is the new web user interface. I used to use forge / reforge, but I guess neo is the most up to date one now. Is that right?
>>8840493Yes, the only "new" feature is the anima support tho
>>8840493Pancho is still updating but I jumped ship to Neo a while ago. Neo does cut a lot of crap out that was outdated or was just not useful, but you also have a much more limited Sampler/Scheduler selection.It also supports Anima if that is something you want to try.
>>8840495>but you also have a much more limited Sampler/Scheduler selectionYou really don't since the anima update.
>>8840494>>8840495Thank you but never heard of pancho. Is it different like comfy, or it more like a different take on how to optimize automaitc1111? Also is Anima the new standard like illustrious was?
>>8840499>or it more like a different take on how to optimize automaitc1111?It's quite literally a fork of automatic1111>Anima the new standard like illustrious was?It's nothing more than an interesting and promising model at the moment since what we have so far is a mere test bake but assuming they work futher into it, it can become the new best anime model for local, yes
>>8840499>Also is Anima the new standard like illustrious was?It's very promising and I've personally switched over already. very much looking forward to the full version.
>>8840503>It's quite literally a fork of automatic1111Ah, the more you know!> It's nothing more than an interesting and promising model at the moment since what we have so far is a mere test bake but assuming they work futher into it, it can become the new best anime model for local, yes>>8840507 (as well)That's good to know. I'll keep a look out for it. I'm currently using WAI-illustrious v 16 so I was just curious if there were was anything better I could use
how the fuck hassaku 1.3 knows the sparkle_(honkai:_star_rail) tag, but 3.4 doesnt??
what command lines are you guys using on forge neo? are you using --xformers --sage --flash --fast-fp16 to speed up generations? With these commands i went from 21 secs each gen to 15 but i'm not sure if it fucks up the anatomy or not
what
>>8840528also is anyone having trouble with the extension infinite image browser? i can't make it work on forge-neo.
>>8840528yes
>>8840526actual good question>sparkle related>>8840528actual bad quest>non sparkle related
>>8840532I just switched to the Desktop version of Infinite Image Browser. The extension worked for me when I used it on NEO a few weeks back, but it was spitting out error messages so I decided to just switch to the desktop version. I also think you are just supposed to pick flash or sage. Not both.
>>8840534How many BBCs can Sparkle take up her ass at the same time?
why do you want to see her poop?
CUNNYYYYYYYYYYYY
>>88405541, she is not greedy
>>8840561ask the gays for some cunny
>>8840569how to get this quality on local?
https://github.com/quickjkee/modulation-guidance wtf is this, it has been added to neo with no explanation
Anima bros
>>8840581Fuck my localchud life
>>8840581FOSS always loses.
kys tranny
>>8840584
>>8840581Another year of glorious SDXL
>>8840581Chenkin snake oil flow won...
>>8840581lol
>>8839918comfyui should donate to whoever makes a better tagged booru
https://litter.catbox.moe/knpqq9tr2t61vyjd.pngNeed foreskin anima lora
>>8840597sorry, canceled
>>88405810/10, apply yourself
>>8840528Tried it just now and it didn't change anything to me anon
>>8840581off yourself, nigga
How many @artists is too many @artists for Anima?I'm using my 12-artist mix from Noob but I'm not liking the results..
>>8840645>How many @artists is too many @artists for Anima?2
>>8840641What GPU do you have? I have an RTX 4070 Ti Super, iirc I think it only works on 3000 series or newer but i'm not sure. Are you sure you’re writing them like this in your webui-user.bat?"set COMMANDLINE_ARGS= --uv --xformers --sage --flash --fast-fp16 ">>8840540I installed all the requirements in the same venv as forge-neo, plus one that was missing, i think it was imageio-ffmpeg, and then it finally let me install it. But it still wasn’t working for some reason. Then I downloaded the desktop app, ran it, closed it, and opened Forge-Neo again, and it magically worked. Weird.>>88406452 or max 3, i'm just waiting for the full release to start autistic mixing artists.
>>8840597metadata for that handsome hetero oji-san?
>>8840646>>8840648Ty, I downloaded some sample gens that looked good and they indeed have 2 artists.
>>88406542 is already too manyanima cannot mix artists
>>8840645I've been having success with some of my old noob mixes by limiting it to 3 and 4. Anything above that, I try to trim out as I'm realizing a lot of those extra low weight artists were just placebo MOST of the time. That said, I prompt schedule 2 artists while the other 1 or 2 I weight accordingly based on how much I want the style to influence.Also, as much as I prefer 2m sde, I'm finding euler a to be a lot more flexible with mixes.
also what do you use to prompt schedule on comfyui? on forge neo it works as intended.
>>8840661I think there's custom nodes for that but I haven't looked yet, I'd love to know as well.
oh my fucking god bruh...https://civitai.com/models/2438778/
>>8840690p*interly bros raise up!
>>8840690kurumuz found dead in istanbul
>>8840690This is why local will never die. So many inspiring creators training whole new checkpoints every day
>>8840690god, i love the tasteful hint of realism so fucking much
>>8840702implying that only nai and anima exist
Detailed bright background for Anima? Prompting "bedroom", but it's always low details and dark.
the fact anima can't blend artists makes me wanna kms...
Fapped four times just today. It's been a while since that happened. Anima is fucking amazing, bros. Being able to choose which girl has the penis makes things so much easier. Being able to prompt who is doing what, who is wearing what, what expression each character has, and who’s fucking who. Plus the updated dataset. This shit is a godsend. I can't wait for the full release. This is what naifags must have felt like when nai v4 dropped, I guess. If only artist mixing actually worked, hopefully they fix it in the full release.
>>8840718I don't want my girls to have penises.
>>8840718>This is what naifags must have felt like when nai v4 dropped,localkeks always were sour-grapes seethers since day 1
>>8840718Based futa enjoyer
>>8840716it can
>>8840724how
>>8840718and that's just preview>>8840726@artist1, @artist2
>>8840728that works horribly
I'm just kinda happy after lurking here through all these dead months. Hopefully this place comes back to life once again. >>8840719Don't worry anon. I also fap to loli x ugly bastard and otokonokos x ugly bastards. >>8840728Artist mixing kinda works that way but it's just not the same as on noob/illus. Maybe I should try prompt scheduling but I don't want to start mixing autistically until the full release
>>8840736Based happy all-rounder enjoyer
>>8840731it's local after all. just train a lora and weigh them differently
>>8840741fuck off
>>8840736it's impossible for it to come back to life now that it has "people" like you using it
>>8840743well thank god we still got that tasty sexy black penis spammer nigger!
>>8840744Faggot garbage isn't any better than tbp spam.
>>8840745because you said so?
>>8840744Hey, i was about to post :D
>>8840747At least add some zest to your pics. Upscale them, make the tbp so juicy that even the haters wouldn't be able to resist. Shit you're posting is so fucking boring.
>>8840690>CATGODDESS CODED/hgg2/, faggot. We're not asking anymore.
>>8840749There are haters here? lmao, some of you need to grow a pair.
>>8840749nigger, it's shit spam, literally
>>8840736>otokonokos x ugly bastardsHow are you coping with having to change over to trap_\(formerly otoko_no_ko\) with anima?
>>8840754Based realizer.
>>8840754so ani was right about him all along. tdrussell is a total fraud. comfy should have never betrayed ani
>trani is seething on /h/ after failing to bait anyone into teaming up with himpathetic really
this is so sadpost more black turds
I gave the wan VAE a try, it's actually sharper but noisier as consequence it's especially apparent in the facehttps://files.catbox.moe/52mplb.mp4
>>8840581LMAAAAAAAOOOOOO
>>8840785it's fake
nai wins again
why is single artist replication on nai 4.5 so weak compared to anima
>>8840803Try Kobuichi and then Koku on NAIv4.5 and on Anima and tell me if that's a bad thing.
>>8840648Gave it a further test and it does seem to give me an increase in 10~13 seconds but it comes at the cost of overall image quality, didn't notice any decrease in anatomy though
>>8840780why is this a video
>>8840834because I couldn't figure out a good way to show the difference
is anima supposed to be that slow?
>>8840856about 3 times slower than sdxl
has anyone tried training a decensorer on kleinseems like the 9b can kinda remove black bars already, but i wonder how much of a struggle it would be to make it learn genitals from all the safetycuckery
who the fuck uses klein for img sloppa
>>8840755Ehh, I mean, the tag is shorter now, so that's better, I guess? But I'm using the tagcomplete extension, so it doesn't really matter. Though you really, really need to use "trap" now, because the other day I was prompting some trap x ugly bastard. I used "otoko no ko" instead of "trap", like I always used to, and it turned the trap into a bald ugly man, lmao. We used to call them traps back in le good old days, so I think it’s an improvement. I always call them that outside of prompting anyways. >>8840826You're right, and I'm pretty sure the one most responsible for that is --fast-fp16. Remove that one and keep the others. Though, I think those flags are mainly for installing the packages, once they're installed, you can remove them and just keep --uv.The --fast-fp16 flag helps, but it only speeds up generations by 1-2 seconds, so after removing it I'm at around 16 seconds per image now. Also, now that I remember, I tested that flag on noob some months ago, and yeah, it messed up the overall image quality and the hands.
>>8840883Gay. Dicks are touching.
>>8840883how do I find an autistic girl to do this for me and a bro?
>>8840883how do I find an straight bro to do this with me and without a girl?
>>8840879Definitely not three times slower with my RTX 5070 ti
>>8840856I used to get around 50~56 seconds on SDXL, now I take around a min and 15~20 seconds on Anima. That said, there are caveats.First, I actually run 20 steps and it gives me the same time as before, and I haven't noticed any significant downgrade for this to be problematic. Second, even going with 30 steps, the less shitty VAE simply solves the issue of crushed faces and other details, so I don't actually run adetailer anymore. Most of the details in the gen is already great by default if you're not an autist running 4k.
>>8840883I want to do this with a cute blonde, blue-eyed white twink
hgg2 will have no gays, just saying...
faggotry is /hgg/ culture
>>8840969nah the fag only came here once the golden age was over
>>8840972so, the golden era of /hgg/ was over when noodlekek created /hgg/?
>>8840973yes
>>8840973explain what's wrong with noodlechad without sounding like a salty fag that came crawling to his thread.
>>8840973the golden era ended in 2023
the churl era starts now
beanerkek has always been more apropos.
Did the thread get even worse wtf
the what and what
>>8841001basedwe love lots of juicy animal penises here
>>8840581Well shit.
>trained at 512x512>less params than sdxl>slower than sdxllocalkeks... what are you even doing??
>>8840294dpm++ 2m sde, sgm uniform or beta_1_1, 4
why do some gens seem to carry elements from the previous gen on forge neo? is this some setting or what?
>>8841047I think it's inherently from SDXL models, it happen to me a little too much times even back then while using NAI v3, some parts of past prompts sticked around for a little
how do i hold back the urge to use turbo loras?
>>8841047>>8841054Fuck, I thought I was the only one. I know some schizo anon was saying it had to do with some 'last gen' influencing thing before he was laughed out of here but I do see it occasionally. Another anon also mentioned a lack of seed variance in anima which seems more plausible.
>>8841060I was genning a guy with a coat, and the coat had off-white lapelsI removed that from the prompt and hit gen again and it gave me a multiple view with the proper version and the one with the off-white lapelsWeird as fuck
I'm trying to find info on how to consistently place characters. I already know regional prompting but that doesn't really work when the characters are supposed to be overlapping, at least as far as I'm able to manage. I saw mention of using inpainting to fix it, and a bot told me to use custom masks with "attention couple" scripts, but that seems like it would only work if I'm using a controlnet and draw over an image with the pose, but both seem like bandaids over being able to control the initial picture. Also, I'm using Comfy if it matters
Anima is REALLY FUCKING ANNOYING with the multiple views nglYou prompt a character standing up for character designing and it insists on doing two
>>8841073holy skill issue
ANIMA IS SAVEDhttps://civitai.com/models/2441085/shutenanon-ai-style-anima-preview
>>8841075nai at home
>>8841074what blud >>8841073 YAPPIN bout?
/hggsoontobe2/ IS following Russell's advice, right?https://huggingface.co/circlestone-labs/Anima/discussions/60#6998b3f592a56ca2caee79fd
>>8841084yeah sd-scripts DOES train it by default
I want to make a lora based on isometric/game cg style but some images are at really low res ~512x, should I keep it as is or upscale it with something? if I do upscale it which one is best?
bros, anima can do that scene of a girl doing a handjob to your pov while on a tv in the background she's getting fuck by someone else. and is quite easy to do
glad you finally catch up, chattook you a little while
>>8841107it can also do girl licking pigs dirty anus
>>8841107no it cant
>>8841110HOW
>>8841110just tried, actually works lol
waow so butiful, is this really ia?
I'll never forgive that fucking Toyota Miyahacki, bros. Never..https://files.catbox.moe/dhl049.png
>>8841116>>>/hdg/ is that way, buddy
BREAKING: nai V5 already in the works and promises to DESTROY local
nothing to destroy when local is still playing catch up with v3
Anima makes some pretty cool steam locomotives
>>8841249Can you try to do this again but using a 1472x704 res?
>>8841250Why that res specifically
>>8841262So you can see the locomotive better, thanks
>>8841249>>8841262Reminds me of that one comic where some guy fucks his mentally retarded daughter. And then some guy in the comments section depressed everyone with trivia about the poor conditions of retards in Japan.
>>8841263Ah, of courseHave the box, for good measurehttps://files.catbox.moe/kfegx0.png>>8841264kek
>>8841267I feel sad now. Be nice to your autist OC.
>>8841073multiple views in negatives?If you're on nai you can also>-1::multiple views ::in positives, which should really kill it
>>8841269She is a good girl
Heyo sirs, Anima + bibici = fun
Mildly obscure waifus: Kaneda Maria
>>8841284canon image btw
>>8841294waifu doesn't mean what you think it means
>>8841294squeezable tits and ass
>>8841169Absolutely kinoConsidering rewatching Bleach just for her
https://files.catbox.moe/30j2xa.png
>>8841320why the catbox? also noob based shitmixes can do her with mere prompt btw
anima is really great at learning styles, but my bakes still fuck up anatomy noticeably on messier styles.
>>>/g/108303964Rouwei T5Gemma-adapter trained for ChenkinNoob-XL-v0.2-Rectified-Flow
>>8841363can it into porn thougheverbeitand can it into porn better than anima albethoughie
>>8841363sounds interestinggonna let someone else do the experimenting first though
>>8841363what does this mean and how does this help me to improve my current sloppa content
>>8841325I bet she can tell you the most obscure and fascinating train factoids.
>>8841363Can people stop doing this frankenstein shit with SDXL already
>>8841400it's just another iteration of bluvoll's rectal flow, pay it no mind
I know this is the most unhelpful thread on this website but does anyone know any artists for that new Yugioh/TCG artstyle vibe?
>>8841439yes, @bumbumguloso on anima
>>8841363... anima full release when?
trained a character lora for animahttps://mega.nz/folder/47Yj3ZIS#klaoBwVZI_u5DbjmCjkqRQ/folder/N64jiaxblora_easy_training_scripts toml i used is in there if you want it, also works for style loras
>>8841486Mind if you share even just a subsample of the dataset? Trying to see what the captions should look like.
>>8841492disnt plan on uploading the dataset because my captions are dogshit because i was lazy but they're all set up like this. you could probably go without the plain language caption at the end and be fine"fuzuki miki, 1girl, blonde hair, green eyes, solo, multicolored hair, green hair, mole, mole under eye, smile, blush, open mouth, streaked hair, short hair, fang, braid, indoors, hair ornament, triangle hair ornament, dress, choker, window, upper body, pink hair, looking to the side, green flower, hairclip, bow choker, chair, collarbone, ribbon, medium bangs, tree, buttons, white dress, detached sleeves, ribbon choker, pale skin, :d, flower, double-breasted, flat chest, neck ribbon Fuzuki Miki, a young woman with blonde hair streaked with green and pink, gazing to the side with a bright, open-mouthed smile and rosy cheeks. She has captivating green eyes and a small, endearing mole beneath one eye, a tiny fang visible as she smiles. Her short hair is styled with medium bangs and a delicate braid, adorned with a triangle hair ornament and a charming green flower hairclip. She wears a white dress with detached sleeves and double-breasted button detailing, layered over a lilac vest, and a black ribbon choker tied in a bow around her neck. A faint blush dusts her pale skin, accentuating her collarbone. She's seated indoors near a window displaying a blurry view of trees, the gentle light casting soft shadows. A green flower is subtly placed near her shoulder."
>>8841493all those caption tags feel like bloat to me when training or even prompting, does it even help?
>>8841499yeah thats why i said in the readme that i probably wouldn't use an autocaptioner the next time and only use plain text for specific things that can't be represented in tags like text in the image. describing the entire image again like that is rarely useful when actually genning.
>vtumor lorasi sleep
>>8841504sorry it's not a "smegma on black juicy penis" lora
>>8841508that's /hdg/, retard we like based stuff here, like churl and animal penises
>>8841511this, but unironically
>>8841486>lora_easy_training_scriptsWait, it actually works with Anima? Since the author had not updated it in ages, I thought it would not. Still, thank you very much.
>>8841528eat my shit, brownie
9001 years old.
>>8841068Attention couple is just a different UI over regional prompting, only potential difference here is if your regional prompter is set to latent mode and not the newer attention mode which couple uses by default.Depends on the degree of control you want and how much effort you want to put into it. There are plugins for most 3D software that can produce an OpenPose skeleton for controlnet. You can just do a rough sketch placing people where they are supposed to be, or generate a reference image with the overly-smart SFW models like banana or grok where you can just tell it "a little more to the left" until it gets it right. You can do a rough gen then inpaint to change the position of on arm or a leg. Use Anima for a base gen using natural language, then img2img with SDXL if you can't get the style you want. etc.Give a concrete example if you want, I might give it a try.
>>8841276Seeing as he mentioned Anima by name, he's probably not on NAI.Note it does respond to natural language in negatives too, if you somehow find multiple_views not sufficient.
>>8841534Does attention couple work on anima btw? Shit just keeps bleeding across characters despite the NL capabilities of the model.
>>8841398she likes planes too
>>8841534The difficult pose would be overlapping bodies. An example I can think of would be a guy behind a girl with thighsex, with the girl still facing the viewer. I built a workflow that at least lets me make a mask I still haven't really messed around with Anima, so I should probably try that. I've been busy trying to get inpainting in comfy to be half decent, but I'm giving up and just setting up Krita tomorrow probably
>>8841541>me make a maskI got distracted and forgot to finish my thought. make a mask to choose where to place people* but it doesn't really let me force overlapping people. OpenPose might work for that though, just more effort than I thought it would be when I started
>>8841541Surely just prompting for thighsex and facing viewer would be enough. I don't see why you'd need regional prompting for something so simple
>>8841543It was just an example off the top of my head. If you want something more complicated, pretzel mating press against a wall with it being an xray from inside the wall
>>8841541Inpainting in Comfy is done using MaskDetailer from ImpactPack which matches the behavior of "soft inpainting" in webUI/Forge, only the settings are named differently. I can write up a mapping if someone wants it, maybe leave it in a rentry.I find regional prompts to be pretty bad at determining composition. Feels like it's mostly set up by the overall prompt, and then regional attaches its features to whatever character ends up in that area.>>8841539The underlying tech works on any architecture, though your UI implementation might be coded in a restricted way. Here's Comfy: https://litter.catbox.moe/kj0ymdlh2auh8g7z.pngCharacter knowledge is rough as always, but the regions work.
>>8841550I'd save any kind of write up if you have the time for it. I was struggling with getting it to be smooth or even work in Comfy compared to forge. Comfy needed 20 different nodes to set up a mask, soften/blur it multiple times, etc. and that still ended up with a hard seam.
>>8841544>against a wall with it being an xray from inside the wallI can't really conceptualize how this would lookdo you mean pressing against the fourth wall? as in the wall isn't actually visible?
>>8841575and what does a pretzel mating press look like against a wallcan you post an example kek
>>8841551Added onto the Forge guide, for reference. https://files.catbox.moe/zy2xlr.jpgConsider replacing the one in the OP >>8839919>Inpaint: https://files.catbox.moe/fbzsxb.jpgHere's a workflow where I added pubic hair https://files.catbox.moe/qgsgij.pngUse the Load image node to load an image, then right-click the Load Image node and "Open in Mask Editor" to draw your inpaint mask.
>>8841527https://github.com/67372a/LoRA_Easy_Training_Scriptsyou need this fork which is where all the main development is now
Just release the 1024x version, bro...
>>8841666wew yeah, 0.6b is good for nowthe resolution is not
>>8841588Nice, appreciated anon. I'll try this out later today
>>8841666That's actually terrific news if the TE really is the bottleneck, which going by intuition, it is. I mean it's 600 fucking million parameters, "L"LMs like that struggle to generate coherent sentences, how can they be expected to NOT be the bottleneck when embedded as the TE for an imagen model?
>>8841666monkaWeird
>>8841716>which going by intuition, it islol
0.6 BILLION CUNNY
>>8841666>anima preview releases>bunch of people complain about the small text encoder>some even call it a "major mistake" and question the competency of the dev>dev runs an experiment to switch to qwen3.5 2b>nothing even decided yet, simply gathering experimental data>NOOOO STOP you have use keep using the lighter text encoder THINK OF THE VRAMLETSMan I'm glad the dev is just ignoring everyone and doing his own thing. It's not just this, the furries are incessant about adding e621 to the dataset, a lot of people are seething about the non-anime datasets it's being trained on, etc.Just fucking let him cook.
>>8841749brown hands typed this post
Bro is cooking but we have been starving for years :skull: :skull:
>>8841749>THINK OF THE VRAMLETSThat's not much of an argument when>a) 0.6B takes basically no VRAM at all and 2B still has very democratic VRAM requirements>b) LLMs this small can be offloaded to RAM with no significant performance overhead
>>8841751brown hands are the ones trying to backseat every aspect of the anima training process
>>8841749Too bad his retarded choice to use an LLM adapter is making the model shit for lora training
>>8841759>retarded choice to use an LLM adapterHow else are you supposed to achieve NL comprehension? CLIP and T5 are clearly incapable of it. All SOTA imagen models are using LLMs as their TEs so it makes sense to follow suit.
>>8841504i've got a 2.5d civitai animu lora for anima too if you would perfer that
>>8841760why are you even commenting on anima's architecture when you clearly don't understand it
>>8841759>LLM adapter hurts lora trainingdo you have any proof at all for this or are you just making shit up
>>8841762please redeem the lora sir
>>8841769lora redeemedhttps://mega.nz/folder/47Yj3ZIS#klaoBwVZI_u5DbjmCjkqRQ/folder/QqoV0CRI
>>8841770thank you sir
cute but also funny?
>Anima on Forge Neo>gen one image with Forge Couple>gen a second image without forge couple using the same prompt and seed>the images are identicalSo Forge Couple doesn't work on Anima. Why though? Shouldn't it be architecture-agnostic?Couple:https://files.catbox.moe/n1qyro.pngNo couple:https://files.catbox.moe/iwu4au.png
>thread is good>fag arrives>thread quality declines>naigger arrives>thread quality declines>blacked cuck arrives>thread quality declines>pedo arrives>thread quality declinescurious
>>8841779fweehh?
>>8841779>Shouldn't it be architecture-agnostic?The tech is, the implementation in the extension probably not.
>>8841781bake hgg2 already
>>8841781so when is the thread actually good?
>>8841797when there's no posts
>>8841779yikes... local is so far behind. Just switch to Pixai Tsubaki.2, the new DiT model by AngelBottomless available on Pixai. Blows even NovelAI out of the water!
Tried to install wd14-tagger and it destroyed my forge-neo, i couldn't even open it. According to chatgpt, it was because some incompatibility and it suggested my to delete tensorflow (it's a requirement for wd14-tagger) from the venv forge-neo is in. So i did, and it fixed it. Just saying in case anyone come across this error.
>>8841822Adding to this, I just found a solution. I haven't tried it yet tho. https://github.com/picobyte/stable-diffusion-webui-wd14-tagger/issues/122
>>8841825I was about to try this solution, but I'm a lazyfag so I installed this fork instead: https://github.com/otacoo/stable-diffusion-webui-wd14-tagger Everything is working relatively fine for now.
>>8841819>Pixai Tsubaki.2I don't use this site, but what are the chances this is literally just finetuned Anima? Timing is a bit suspicious.
>>8841838it's way better than anima. if it actually is a finetune of anima that's more of an insult to anima than anything, that they managed to pull off a finetune that improves the comprehension and aesthetic by 5x in under a month.
>>8841841surely you can post your gens as proof then
>>8841842i don't gen, just shitpost
>>8841819it's actually crap and doesn't know artist tags
>>8841847>artist tagswai base style is all you need
>>8841828So this gets tagger back on Neo? Because I too had this >>8841822 issue when I first flipped over to Neo and unlike anon didn't ask chatgpt so had a nightmare reverting everything back to working condition. I've since just ran taggui individually but can't say I miss having everything under one roof.
>>8841825>>8841828So is it actually working?>>8841858I just reinstalled it when it happened to me lol
>>8841849this but unironically
Dudidam
>>8841865>>8841858Yeah, it's working, I've been prompting for hours and everything seems fine. Just use the fork I posted.
>>8841865>I just reinstalled it when it happened to me lolIf only it was that simple, anon. I couldn't get this nigga to run normally unless I went through some uv method an anon shared here. Pancho played us for the long cuck.>>8841874Nice. Thanks.
>>8841877You need to at least read the instructions when you install anything anon. Remember that.
local is savedhttps://civitai.com/models/2402121/imagine how brown the hands are
>>8841900For me? The pussy bulge, sirs.
>>8841900>animawhy does every local model wind up looking like that???
>>8841916brown hands itch for that sexy look.giving mongoloids the ability to modify models was a mistake
>>8841900>>8841917>grey with slight warm morning color tint>butiiful 2.5d sexy realism>same compositionwhat are their datasets even? highest rated ponyslop of 2024?
All Disney Princesses from Ralph Breaks the Internet Anima lora when?
>>8841928This + incase + afrobull + expressiveH please
>>8841874can you post her wearing that
https://pixai.art/en/tsubaki-2/guideprompt comprehension seems pretty good but I'm not sure that matters if the output still looks like shit
>>8841948kino
>>8841943kek this piece of shit was trained on a bunch of niji gens wasnt it
>>8841951I tried reading about it and no ones talking about it anywhere aside from some on r*ddit and supposedly has a really strong base style, so yeah, it's slop through and through
>>8841970that's a boy
>>8841972it's 1girl, flat_chest (not loli)
>>8841973pedo
Has anyone trained one of those slider loras for anima yet?
>>8841828Didn't work for me, but thankfully it didn't nuke my forge neo kek>>8841984Never understood how those worked
>>8841998based on the lora strength you get more or less of an effect. Dunno how to train those
https://civitai.com/models/2447159/skinsuit-for-anima-or-2333333
>>8842009This is on the same level of royally fucked up as gore, scat and NTR ngl.
Man, the lack of autotagger is the only thing stopping me from migrating to forge neo permanently. And I really don't want to install taggui.>>8842004Yes, it's a slider lora, the way it works on practice isn't a surprise. I mean to say I never understood how it worked under the hood, as it seems to be different from regular concept loras.
>>8841998>Never understood how those workedShame. I wanted to make one to increase/decrease the mesh size of fishnet fabric, but I'm clueless with concept loras, nevermind slider ones
>>8842013agreed kek
I like NTS
>>8842013what has to go wrong in someone childhood that cute anime girls with big ol' saggers don't do it for you anymore?
>>8842015Personally I love slider loras, they give you great flexibility and basically don't mess with image quality and style even when made by civitniggersA "nose size" slider lora will always be better than a "small/big nose" concept lora imo
Any tips on genning characters stepping on the legs of a squatting character? Also horse penor.
>>8842032like this, or what? kekdoubtful
>>8842035I suck at explaining lmao. Like this but for squatting, while the guy stepping on is standing/rubbing penis on face or breasts. For some reason it's also hard to do even when the character is on her back like pic related too.
>>8842041I still don't get itso it's probably too confusing for AI too lel
>>8842047Fuck man. Thanks for trying though.
churl'd cunny
>>8842041I still didn't get what you want. But as a general rule, if you cannot find multiple reference images on Danbooru/e621 and have to resort to text descriptions then no, the model cannot do it without a concept lora, or possibly controlnet.
>>8842118I'll try another one but it's just this and similar but with the guy literally lifting his leg(hands free) and stepping on the female.
>>8842122ohhhh for the guy squattingyeah now it makes sense
>>8842035I do this and look like this
>>8841970can you run anima in stable diffusion now?
I'm sadAnima can't do [spoiler]scat[/spoiler]
>>8842179that's why we need e621 dataset
this, but unironically
キモい
>>8842032>>8841874plz box
>old>reddit>stablediffusionUNC hours on /hgg2pending/
the fuck
why is every thread that has to do with this hobby so autistic :Pepesadge:
:P
cunny
does
get me hard
anima is so goated, impossible to do this on SDXL using adequate tags, controlnet and proper inpaint methods
Been using Anima a shit ton the past couple weeks, here are some things that work best for me and I'd like others to weigh in on what they have tried. Some of this might just be placebo/etc, could be retarded, but figured I'd share anyway.> Euler A with Beta scheduler and 4 CFG works really good> Prompt formatting matters a lot, you want to make sure there's a comma and space after every word. For certain character names or artist styles, you have to use / if they're in ( ) or it'll be fucked. If you prompt for Asura it needs to be @asura \(asurauser\) as an example. The suggested tag order on the Anima HF isn't a strict rule, but generally following it seemed to help. For NL, it's important to make sure there's multiple sentences or it'll degrade quality.> It is possible to upscale images using multidiffusion without introducing artifacting (I'm doing 2x upscales on all my images). Assuming you're using forge neo by Haoming02, Activate "Mixture of Diffusers" and set the tile width/tile height to the resolution of the original image. For 50 steps I use a denoise value of 0.15 and do the upscaling in img2img along with adetailer pass after (which also does work!) You do not need to have an img2img upscaler selected either> If you can't get a certain low count character or artist tag to work, using a shit of supplemental tags can get it to go through (for art styles, you can use NL to describe what the style looks like)- Finetunes are all snake oil for now or WAI slop, but loras can work great and I don't really notice the knowledge loss people are talking about (although I think the earlier loras people were training had this problem, there's discussion on HF about how to best train them).> The negative matters A LOT. My neg is super schizo and it noticeably improved images, you can also use NL in the neg. When it comes to artist signatures/etc, the longer the prompt the less of a chance they appear in my experience.
>>8842262Care to post the neg?
>>8842262>denoise value of 0.15Isn't that really low
>>8842267that's the point
>>8842268why? mixture of diffusers usually doesn't start to hallucinate full prompt in tiles until ~0.45 denoise, why so low? you get artifacts with higher denoise? imo below 0.3 denoise upscales don't look good.guess I can try it later myself
>>8842262>you have to use / if they're in ( ) or it'll be fuckedImagine actually typing out an artist instead of using tag autocomplete>it's important to make sure there's multiple sentences or it'll degrade qualityYeah, 8b llm would've handled complex sentences without any issues, but for 0.6 you have to dumb it down as much as possible ('The girl on the left is standing with her right hand pointing at the viewer. She has blue eyes and long blonde hair. She is wearing xyz')Euler ancestral is really amazing for anima, it goes along with the majority of schedulers, though I prefer normal. er_sde with sgm uniform also works great.>denoise value of 0.15In my very limited testing esoteric schizo tests 0.24 seems to be an absolute maximum before it starts hallucinating (though I use ultimate sd upscale)
>>8842227how's the character knowledge?
>>8842262>multidiffusionIs it supposed to be THIS slow and massively lag my whole system? I can't even interrupt the generation, I HAVE to close the terminal tab itself.
>>8842275Actually nevermind, this seems to only happen if I use "Resize by". "Resize to", albeit janky, works fine. Weird bug though.
>>8842266Sure, you'll probably need to adjust it depending on styles you use though.worst quality, lowres, scan artifacts, dithering, halftone, screentone, chromatic aberration, vhs artifacts, jpeg artifacts, adversarial noise, film grain, sketch, comic, artistic error, unfinished, retro graphics, lowpoly, blocky graphics, pixel art, cropped, letterboxed, border, bad anatomy, bad hands, extra arms, extra legs, extra digits, missing finger, deformed hands, extra limbs, mutation, deformed, extra eyes, malformed limbs, fused fingers, too many fingers, long neck, cross-eyed, bad proportions, missing arms, missing legs, censored, censor bar, mosaic censoring, signature, watermark, artist name, logo, text, sound effects, web address, speech bubble, muscular female, abs, ribs, crazy eyes, @_@, mismatched pupils.
>>8842272Yeah I do use tag autocomplete, but I know some don't. It's honestly easier to write dumbed down NL prompts, so I don't mind the 0.6. Apparently he is testing more params, so the final version might be higher. I'll have to try normal.
>>8842262Like that? I'll try it for AnimaYume 0.2 too
>>8842273nta but it's quite bad, not being able to use (((weighting))) kill it, it makes it impossible to emphasize the character over the rest of the tags. its going to be another lora-reliant model because even characters with lots of hits aren't fully accurate.
>>8842327what characters are you having issues withI've not run into any troubles
Do you think she flicked her bean after getting cummed on?
>>8842327its an even bigger issue with artist mixing because some artists are just way too strong and there is no way to dial them down
nigger use prompt scheduling, we have been over this already since day 1
>USE THIS SNAKEOIL THAT PERFORMS WORSE THAN SDXL!!!
also anima does have prompt weighting, in comfy at least
>>8842262>Euler A with Beta scheduler and 4 CFG works really goodIt softens the image which can help with hands, feet, etc. But it ruins some art styles. I find euler to be a better trade off. Heun is low key great if you don't mind the longer gen time.>For NL, it's important to make sure there's multiple sentences or it'll degrade quality.Dunno about that but being clear, concise and writing like you would using a lobotomized llm, works wonders.>It is possible to upscale images using multidiffusionStill on 1.4x hiresfix. I no longer feel the need to upscale bigly like I would on 4ch vae.>not prompt schedulingkekerina of time
>>8842339juiced up
nothere has to be some rule against this
yup, the sanction: you fucking off, faggy
against what
>>8842355piss, probably
you faggots sensitive over anything besides black scat dicks
pissing in uterus is straight from e621, let's not pretend otherwise
oh no!
>>8842273Uhh it is *check notes* very good! , the lack of any furry images in the dataset leads to an amazing range of truly anime style and concepts Anima is the anime model
>>8842339Based and hot. Total pov cowgirl and missionary vanillatard death.
>>>>>>>>>>/slop/
where is the CUNNY
>>8842369already told you to ask the gays for cunny
>>8842371this is the hentai gays general
who do ask for hetero black veiny cock gens?
>>8842375/hdg/
>model with a million devianart monstrosities in the training data>anime model
Can I somehow stop joy caption from starting every description with "A Digital anime-style illustration". I feel like that might mess up training
What's the best local llm for writing prompts? 24gb?
>>8842401Why do you need a llm to write your prompt.
>>8842404Thought it might be fun to not know the exact result I'm going to get while having a little more control over the general direction than with wildcards
Why are mass sloppers like this
>>8842409Atrophied pyneal gland from a life of maluntrition and lack of physical activity.
>>8842338I'll give Euler a try. Aren't you running into artifacting with a 1.4x upscale that's not tiled? I tried and ran into issues, how many steps are you upscaling by?Unrelated, but this is useful - https://github.com/KohakuBlueleaf/z-tipo-extension
Hit me with your strongest ntr/nts gen
>>8842337I am very confused since prompt weighting seems to work on my side and I use Forge Neo
So, umm, what am I supposed to use? Anima isn't finished and everything else is super old...
>>8842497Don't use anything, turn off the computer
>>8842497Pick a pen and use your imagination, NPC.
for me it's a graphic tablet and digital mixed media
>not using animal blood and a good stonengmi
Most unhelpful thread award
/hgg2/ when?
Maybe if some idiots were posting anima gens...
so now that /hdg/ is completely dead I guess I'll ask here
post black penis images please
Do we know what the token limit is for Anima?
Anima is a decent model for late 2024, but the tech has advanced since then so it already feels outdated.
you mean nai?
nai is nsfw sota doe?
We don't share meta in these parts, Rajeesh.
>>8842534black penis images
hello? blacked lolis?
no, sweaty cunny pitpussy however
well where is it
Frankly? trap_\(formerly_otoko_no_ko\) kneepitjobs.
>>8842362That tattoo is almost precise. Not bad at all.
>>8842583Yeah, anima is really great innit *wink wink*
>>8842227How did you boomer prompt this? I'm still struggling to get off the pure tag and pray on rng system.
>>8842587>boomer promptRead it.https://dskjal.com/deeplearning/circlestone-labs-anima
>>8842587Here's the catbox, bro. Be sure to read the metadata thoroughly :^)>https://files.catbox.moe/8ptixc.png
>>8842590>noob shitmixgo to bed grandpa
>>8842587just use chatgpt bro, it's not that complicated. give it your sfw tags, let it turn them into nlp and you just fill the blanks to turn it into nsfw. Then you learn how he does it and replicate it next time you want a similar scene.
>>8842587just... write.
>>8842589Why are they having outside on my windowsill?
we need the e621 dataset so it can figure out where horse legs gohttps://files.catbox.moe/j9vvlb.png
i agree
How do us uncs feel about zoomers pushing the e621 agenda?
>>8842657
stole this from /g/https://huggingface.co/spaces/DraconicDragon/Kaloscope-artist-style-classifierit tries to guess the artist of imagesgood for finding new artists with the same vibe as a given image
>>8842660interesting...
>>8842660that's cool as shitonly thing I'd add is filtering for how many images an artist has on a boorulike only show similar ones with at least 150 images or so
>>8842660>upload pic with anna_anon>max similarity result is novenine with a 2.6% matchSeems pretty worthless.
>>8842668as someone unfamiliar with either artist, other then one being 3d and the other 2d their art styles do look pretty similar
>>8842668maybe next time try uploading an actual artist
>>8842670what if I wanted a new artist with the same vibeas advertised
>>8842671you uploaded a western anime artist and got a different western anime artist in return seems like its working
>>8842581you're going to need to find a boomer prompting god because this is as close as i can gethttps://files.catbox.moe/c2oel4.png
>>8842675just inpaint disembodied penis
>>8842677yeah but then i have to inpaint on comfyui and im not putting that effort in for a fetish i dont have
>>8842589box plz
What a wonderful day, I hope to see many Anima™ gens today (or this week at least) by the amazing posters of this chat (which everyone are very kind and thoughtful persons that greatly contribute to their community in a positive way)
is it true that anima training stopped
>>8842731No, that "screenshot" you probably saw is a blatant fake. The last real news from the author was from a few days ago when he said he was experimenting with a bigger TE, I think it was Qwen 3.5 4B?
>>8842732>4B2B actually but still a pretty massive improvement over just 0.6B, AND the model itself is quite a bit better when comparing equal sizes.
>>8842759Man, having a better VAE compared to SDXL is so noticeable with images like this.
>>8842764For real, the SDXL VAE would've mangled those hands and eyes beyond recognition. And that's just the 512p bake.
frfr? no cap?
but can anima do rusty_trombone with horses
Jazz for your soul.
>look anima can do a girl doing a triple backflip and sucking 3 dicks at once while making preparing sushibut can it make something that looks good?>....
Could any anon help me achieve this art style? I tried training a LoRA, but the results look terrible.https://www.pixiv.net/en/users/116224005
Only nai is capable of producing good images
>>8842783rapi my beloved
>>8842783Isn't that just dikko with a bunch of WAI slop smeared on top
>>8842787>shiny skin= WAIbased blind retard
>>8842787Yeah it looks like dikko style. Thanks.
Everything is WAI, it's the best anime hentai model available after all.
ni howdy
>>88427332B was cancelled btw
>>8842581>>8842675Based knowers.https://files.catbox.moe/4b5f6t.png
>>8842827That seems really odd. Perhaps the actual bottleneck is the NL tagger? I just can't think of another explanation as to why such a big improvement to the LLM TE would result in no noticeable improvement to prompt comprehension.
2B but blacked
>>8842830dude he is saying it right there that it would take too much training time for it to adapt
>>8842830I would argue there probably isn't much room for it to actually improve. The LLM side of the image gen equation probably isn't a super huge part of the equation until you get to the area of making the model a context/vision model(this is what NBP/dall4 are), which stands to reason that there's a middle point where you gain next to nothing by going larger other than making everything slower.
why gay gens are always less gay than the "straight" ones
>>8842828samefag aidsmonkey
>>8842839you're gay
>>8842841Tourist browns need to fuck off back to their deadname /hdg/
>>8842843try posting your gay shit outside catbox and see what happens
>>8842839elaborate on your very ambiguous statement
>>8842845Trying being white in pre #10 and see what happens.
pre 10 uooohhh
>>8842848your uncle couldnt resist your tight white ass and infected with you the gay parasiteSAD
>>8842850golden cunny age hgg..
monkaGiga
>>8842847lots of the images posted here have odd fixation over male focus, like half meter veiny penises, gangbangs, dp, ugly bastard ass being 2/3 of the gen, etc.but faggots keep coping about muh pov cowgirl and >>>/e/
>>8842855Aside of the groid gens (which are pure shitpost), which ones are like that here? lol
>>8842856stop replying to the faggot or you'll catch aids too
>>8842860post your gen, white heterosexual male
https://files.catbox.moe/znks5e.png
>>8842894do you like konata more than kagami
head way too small
even heads decline
>>8842895konata is 142cm, kagami is 157cm.take that to mean what you want.
>>8842904Her father did not feed her correctly. Even for a japanese girl, she is stunted.
>>8842895konata has bigger head than kagami.take that to mean what you want.
>>8842905konata's mother was only 143cm. it's hereditary.
>>8842907she's smarter???
>>8842909yes, oekakischizo is sapiosexual
>>8842894Nice Anima™ gen you have there
>>8842920that's still noob 0.5 eps tho.anima character consistency is still too shit to be usable(not to mention upscaling is still a bit of an inconsistent shitfuck). also I still haven't bothered trying to train on it yet. I will say it's capable of better cat stretch compositions due to being able to actually foreshorten a bit more correctly.https://files.catbox.moe/2ccgys.png
>>8842924You didn't use Anima™?! but it's the newest bestestest pure real anime out there! new == goodest old == baddest
can't tell if it's satire or if it's really him
>still using sdxl
>>8842924anima can't do white background?
>>8842820I should do more gens with that mcdonald's girl
>>8842927does it matter?
it doesn't matter
>>8842948why is the face oddly blurry (i like a little the gen overall but I am curious)
>>8842949it's nai
>>8842949it's anime
nais? there you go sirs! ^ -^
>>8842956me in the middle
>>8842959white background?
>>8842960i'm brown, tho
>>8842962why are you lying
why are armpits so weird, why they have so many acceptable shapes
>>8842956unf, size comparison is amazing
>>8841486Interesting. I have been running a test for multi-subject lora, and at 2200 steps it gets *most* things right, but it clearly is failing at the details. Will train it for longer and report back with my findings.
Juri still owes me uninterested footjobs.
>>8843032anima loras suck at detail because of the 512p training unless you really burn them
>>8843034fruity stud
>>8843032my lora also has problems with some details but i assumed that was because i trained at lowres and lack of representation in the dataset. i trained up to 5k steps but didnt see much improvement past 3k
>>8843018nai is way too good
/hdg/ is finally back
Why I can't create something this erotic? what I am lacking? what I am missing? ;-;>https://files.catbox.moe/204ayl.jpgmy dogshit slop>https://files.catbox.moe/wtdqol.pngAnd yes, showing part of her ass was intentional, perhaps that's what is making it less erotic? I am lost and mad at my own incompetence chat
>>8843078eye profile: first image she's looking down on the viewer while yours is looking up. eye intensity on the first is also just plainly better.blush profile: first has a more natural blush while yours is a bit more like a blush stickermouth profile: first is parted lips and a bit more apprehensive, yours is just a plain smirkhair profile: first has hair that's a bit more disheveled while yours is just cut clean and neat
>>8843078ai lacks intent. the finer lines highlight details which ai completely misses
>>8843078unironically its the lack of soul
>>8843078just add a horse penis
>>8843078you're missing 25$
>>8843080thank you for your honest feedback, I am seeing it now, I'll see what I can dodo you like armpits? no? me neither but I did a gen about that which I think it's not that great but I am going to share it anyway so you can see it and laught with me about my own misery for doing something that I clearly don't understand>https://files.catbox.moe/mbayou.png
>>8843086I as the baka I am upload the actual image and also shared the catbox link of the same image, this clearly reflect the completely mess of my mind on this momentFeel free to point and laugh at me
>>8843086since you're going in on the humiliation fantasy imagine the girls utter disgust and shame in being associated with you and any of your actions as chat laughs at your utter incompetence.she may be there and might be putting out but she's not very happy about it.
>>8843089:Sadge:Second try, I think it's a little better but still not as erotic, background pattern is a little messy but it can be easily fixed so whatever>https://files.catbox.moe/71gi4s.pngBy the way, go and check the real artist because I think his work is really neat (and erotic)>https://x.com/Leutte_37
have some texturehttps://files.catbox.moe/10lvxd.pnghttps://files.catbox.moe/7g7ou8.pnghttps://files.catbox.moe/j9bjwt.pngshitty 5.6mb lora, trained on vpred 1.0 https://files.catbox.moe/l2vtvt.safetensors
>>8843122>sdxllol
>>8843122>black backgroundtruly daring
>>8843124IT BROKE NEW GROUND
>>8843122>open eyes ichika?
>>8843123eps 0.5 is the pinnacle of henti and nothing will ever surpass it
doesn't vpred 1.0 have fried TE and worse anatomy
>>8843130your wai-nsfw-illustrious! sir?
>>8842855>Straight gens>Look inside>Made by losers who hate women>Gay gens>Look inside>Made by faggots who love (usually feminine) men That' it. It boils down to this. All the ugly bastard, groid, rape, group sex gens and so on are all made to degrade women while the fags just post regular shit but gay
why does insecure aids ridden freak keep samefagging?
Sorry sis but the socks have to stay on or it's gay
I find your lack of cunny disturbing
>>8843144>cunnyNo idea why this 50s slang retook popularity.
anime girl but HUMONGOUS ass and thighs
>>8843149check the other thread
>>8843136TRVKE got the chuds seething
>>8843018we need more like this
>>8843157hdg is that way
i thought we didnt kinkshame here
>>8843159What made you think that?
>>8843165all the ugly anima gens
>>8843165images posted
>>8843167how would you kinkshame someone by posting an image, that seems way too high effort
https://files.catbox.moe/rcdohe.pnghttps://files.catbox.moe/r2um4y.pnghttps://files.catbox.moe/3ibf65.pnghttps://files.catbox.moe/atwo1z.png
>>8843176being hellbent about eps 0.5 for over a year and now moving to vpred 1.0 is the funniest contrarian shit i've seen here
>>8843191I've been using vpred for random bullshit go stuff for a while? Textures tend to train better into vpred so if I want to do dumb shit with a dataset with a lot of texturing I'll default to vpred.it's not illegal to use different models for different things.
>>8843194sure, i just find it hilarious also, the new style is pretty nice
is res_2m/beta57 any good or just placebo
>>8843032Trained for 5000 steps. It works better-ish. The details are far more accurate, in particular things like details on their clothing and hair.Also, bishounen tag give some muscle definition to the boy's neck. Color me surprised.
>>8843176why are you a furry
>>88432081. implying you wouldn't fuck that tanuki2. at least it's not horse cocks
>>8843176>bobblehead KagamiAyase pic is based though
>>8843209I wouldn't a tanukiWhy is everyone still kink shaming me for the oldest bestiality trope out there >____>
Are there any artists drawing bellies like this, without making girls landwhales? https://x.com/wmmzkl/status/2012916298888691736/photo/1 https://danbooru.donmai.us/posts/10769008I'm using belly, fat rolls at low weight, but it's inconsistent and tends to fuck torsos proportions.
>>8843220try muffin top
what does rectum flow even do compared to normal noob?
>>8843247to be quite honest almost nothing, if anything it actually does less than your average shitmix since it doesn't play well with most of the cool samplers, the only good part of it it's the updated dataset and the more sketchy crayon eating styles it can do
>>8843240Thanks. Seems to be the same case like with the other tags. But I reminded myself about shennai misha and with those tags at low weight it's fine.
>>8843250updated dataset but weaker knowledge of the older stuff, cause it's based on eps-
>>8843247>rectum flowdid I miss something?
>>8843297you missed a nothing
Any good anima artists finds? I can't prompt
>>8843307asanagi
>>8843307there aren't any
>>8843307umikochannart
>>8843307apostleincasecutesexyrobutts
>>8843307nyantcha
>>8843307kittewxinaellesane-person
>>8843307ZUN
>>8843307tsunakoiuuikobuichi + muririn (they mostly draw together)kokushida kazuhiroiizuki tasukumitsumi misatoomone hokoma agmyunsangohisashiburiduoyuanjunm-da s-tarouhappoubi jinsofthantenwanacaJust some banger hood classics from the top of my head.
>>8843336great style
>>8843338it's Gemuoa scat artist, but that's how it goes. all the best ones are weird in some way
>>8843339cute hags
>>8843329Huh. Looks like Anima was trained on eroge CG. Some of the images I got look like your typical CG.
>>8843343>what is game_cg tag
>>8843343>what is danbooru
What is love?Baby, don't hurt meDon't hurt me no more
>>8843307I'm one of those types that like butful painterly shoujo-esque styles but I swear I can't ever find one artist of the like that does pits/feet/anus properly
>Art is not consistent enough>But the SFX are....?
>>8843355one of the many wonders of the slopmachine
>>8843355her age?
>>8843355they're also not consistent japanese sfx are dogshit on anima because of the 512p training
>>88433570 years, as I literally just prompted her into existence.
>>8843349Piss is disgusting and belongs in /d/
>>8843362who asked?
>>8843362very true oomfi
>>8843362so true xister, trans rights!
is lactation /h/ or /d/
>>8843385/h/
>>8843385We have a milking thread right now with hundreds of posts. Mods do not seem to mind.Not that lactation is all that depraved. Very tame, actually.
is it me or Anima gives very inconsistent gens regarding the artsyle?Sometimes looks good, sometimes looks like someone messed up with your tags. I'm this close to downloading a shitmix because of this.
>>8843406This happened to me with short prompts, you really need to flesh them out with natural language/lots of tags for now.
>>8843406It's a preview model, way way undertrained, it's a cool concept but people just don't understand shit's fucking raw and needs a very good time on the ovenI do wish and hope for its success tho
>>8843406Kek this exact thing happened just now with meLeft in picrel is for comparison, to show that the style was consistent with the one in the middle, then out of nowhere it gives me rightIt's still tagged full body btw
>>8843406You can try this to see if it helps, but seems like snake oil desu - https://civitai.com/models/2385278/animayume
>>8843406Not only you.It's what I have been getting very often save for extremely popular and prolific artists like Asanagi and the like.Very good as a pre-Alpha model though. So I am still looking forward for a subsequent model.
>>8843412I've had this issue far more with big artists than the small onesThe bigger the data, the more it tends to vary a lot from gen to gen, some anon said that you can fix this by describing the style you want in natural language and it does seem to help but I'm too ESL for NL
forgot to post about this, added a nsfw prompt section to my anima artist listinghttps://anima-artists-prompts-957be3.gitgud.site/nsfw.htmli might actually re-do this prompt because im not very happy with it, way too much broken anatomy that probably wouldn't have happened with a different prompt or steps at 30 but i'll deal with it later
>>8843422If you're going to redo it, I'd recommend getting something that shows feet better
>>8843423i was probably going to remove the feet tags entirely, it's mostly irrelevant to what I actually want shown which are the nipples and genitals
>>8843425Pity, but fair
>>8843422what the fuck, why in many cases it look like the penis was badly drawn over the gen using MS paint
>>8843431you need e621
>>8843446the trvke animasissies need to accept
>>8843446everyone smart is saying this
>>8843446https://civitai.com/articles/26762>We tried higher dropout and more aggressive repeat strategies to strengthen character quality, but the model kept degrading. After fixing dropout and repeat, quality was still getting worse, so we spent a long time debugging. We only recently confirmed the main cause: the e621 dataset. Once we removed it, quality improved clearly and we’re back on track.
>>8843455I can't wait for pussy and penis lora
>>8843455you're a fucking retard if you believe a word they say. they're incompetent.>we don't wanna do any work on the dataset so just remove shit and see what sticks
>>8843457Isn't that the modus operandi of most AI tech companies though?
>>8843457>curate e621 dataset to separate what might be damaging from things that are less likely to be damaging>the result is a dataset so small it's not even worth the risk of adding it
>>8843455simply an skill issue admission which is even more embarrassing a year after naiv4.5's release
>>8843455just... tag properly?
>>8843459There is plenty of good furfag art, but most of it are just plain turds. Going through it and separating the chaff from the wheat is a daunting task
>>8843459how is small dataset an issue? artists with 80 posts work. same would apply here. they're just taking lazy shortcuts that will inevitably make their model just another animagine/hassaku failbake
>>8843461>j-just tag it bro!>meanwhile e621 uses a slightly different tagging system/nomenclature to *booru>this kind of fucks up the TE>the parts that are the same then bleed even if you try and set a master tag for the full datasetthere is no real winning here.>>8843462most of what's worthwhile gets added to *booru anyway. and the anima dataset is gelbooru, which has significantly lower standards/is much easier to add things to.>>8843465you have two datasets.one is known working. the other is known to make everything worse. You can put in a lot of time and effort to cut down the dataset known to make everything worse to 1/10th the size(and that might be being very generous on estimation). But it's still from a dataset known to make everything worse. Are you really going to waste the time to cut it down and then risk even more time and compute to see if it would be "okay?" Or are you going to just drop the entire thing and stick to the dataset that's known to work fine.
>>8843469>>8843456
gang i have started using anima and im wondering if the only way to upscale is to use sd ultimate upscaled? i tried the previous hires fix where u load a upscale model then put a ksampler after, the upscale quality on ultimate upscale isn't as good as the hires fix or whatever its calledhere's a test using the hires method, it does burn the image but puts better details and enhances the already established ones than ultimate sd upscale
>>8843469you have the mindset of a 2022 sloptuner but that's also the case for pretty much every single person involved with anime nsfw models so i'm not blaming you
>>8843469what the fuck are you saying jessie, what kind of different tagging system are you talking about? the boorus and the furries share many tags for the same concept which is a GOOD thing, sure they have more specialized tags for some others but that's also a good thingWTF
>>8843471is this just krekkov or which other artists are mixed
>>8843474its @eigaka, @riz, @krekkovi like the colours and how the skin connects they look like delicious and sweaty cake
>>8843455>chenkinnoobMore like heckin noob lmaoo
>>8843475nice thanks
holy /aco/
>>8843422Some of these artists should be in jail
>>8843480i know, i know. western-style art is very scary but artists shouldnt be persecuted just for their artstyle. even if that artist is fellatrix.
if it doesn't look like cartoon network it's tasteful aco
Dexter's Laboratory and PPG was peak ngl
for porn?
Their age?
fictional
>>8843422>>8843423I use this pose to test out anatomy and shit, it's really useful>cowboy shot and close-up of a woman's feet, kneeling on a white bed on the left, standing on the right, armpits, sweaty armpits, arms behind head, three quarte view, feet soles facing the viewer, close-up of feet and lower body on the left, with her head out of frame, cowboy shot on the right, multiple views, ass, thighs, seiza, hands on own ass, ass support, anus, pussy, legs, soles, foot focus, detailed foot soles, light frown, blush, sweat, steaming body, dark background, uncensoredThen I just describe the girl below, in this case it's nami
How do you tag for a style dataset for Anima? I assume that merely adding relevant tags is not enough.
>>8843495fir the style lora i trained, i just added an activation artist tag and then autotagged like usuali would also recommend removing any tags that should apply to the style by default like "realistic" or "sketch"
>>8843522Noted. Thanks.
>>8843410No, it's not snake oil, for anatomy at least.
>>8843410it's not snake oil, it's slop
>>8843422your safety tags? uncensored?
>>8843545i intentionally didn't add uncensored, i did forget to add explicit though.
>>8843546nta but explicit and safe are snake oil. you should use uncensored
>>8843547i wanted to keep uncensored off for artists like @ame \(amechan17391739\) who are by default censored but adding uncensored could improve the genital quality for artists that dont draw porn. Really just depends on how much of a focus on accuracy to the original artist I want.
anima 102d custom when?
>>8843552not any time soon sadly
https://files.catbox.moe/9n3qwd.pnghttps://files.catbox.moe/ea7f8y.pnghttps://files.catbox.moe/nxtmxw.png
BREAKbros... 2026 is our year. I can feel it.
WAInima when
Does anyone remember that one project an anon shared that cleanses generated image from the nonsense bytes added by AI? Did that go anywhere?>>8843561That black haired OC of yours looks mean.
>>8843567it's where it's always been.https://www.mediafire.com/file/vipr23exc5htmntmay require fiddling with settings on a per picture basis for best results since it can squash gradients.
>>8843565there are already anima "so-called" finetunes using ai-generated slop as dataset, you need more?
>>8843567>nonsense bytescare to elaborate?
>>8843570The reason why it is a bad idea to train based on your own generated images is because of the clumps of data that the AI leave behind. That other anon explained it far elocuently than me a few months ago.Lets say for example that you made a shitmix of artists with your loras or basemodel, and want to reproduce that same style on some other model and not rely on the whims of the RNGods. You could either remove that manually, or use the project this >>8843568 anon provided.I have made a couple of loras that do use cleansed AI generated images, in particular for outfits that I do not want to spend hours inpainting when I can just use a lora for it. It is quite cannibalistic, but it works.
>>8843571so, it apparently blurs the image and applies some filter.
>>8843569If it doesn't have WAI in its name he can't gen his WAIfu with it
>>8843572Pretty much. But it gets rid of the extra nonsense in the image. You may notice that the file size is smaller now.
>>8843572blurring is the exact opposite of what it does.it's closer to a posterizing it. it minimizes the color space and removes latent diffusion noise.
>>8843572>>8843574>>8843575So why not just gen with Kohaku LoNyu Yog and cut out the middle man?
>>8843576>why not use a shit sampler no one cares about or uses that takes twice the time to render instead of using a super simple script that only takes a few seconds to run good question
>>8843575dunno bro
>>8843577>why not use a shit sampler no one cares about or usesThat's because it makes shit look like this trash >>8843571 so you tell me.
oh boy it's combative schizo hours
>why can't they just agree with me!!!
>>8843582let's see your Kohaku LoNyu Yog gens
>>8843571Seems like just passing the image through lanczos would already remove random noise, even without changing the resolutionBut it is a good point in that you shouldn't be using gens as input untouched.
>>8843583I thought that was all hours.
head size decline...
>https://civitai.com/models/2451509/brittle-ai-style-anima-previewah yessir, it's genning time
why would you want your style to be brittle
I'm finna tryna smugpost, lil blud.
pretty cool that throat x-rays work on animai wonder what else i can get to work. paizuri x-ray?
>>8843615Damn, nice Rika.
churlbros.. our response?
Should I gen?
nah, just shitpost about the gens of others
anima is ugly and its never getting an updatei cant afford 25 dollarsits been 3 years i dont want to use sdxl anymore
>>8843655>anima is uglyjust train a nyalia lora
just use thishttps://civitai.com/models/2393785/ntrmix-or-style
>>8843659i came
>i cant afford 25 dollarswhat causes this?
global poverty
>>8843655gen realism, you got z-image and chromachroma2 on the way
lol
>>8843655>anima is ugly>i cant afford 25 dollarsanon you'd have to fight 25 dollars against ugly bias even harder. at least with anima you have loracope
>lorasno one actually uses these right? a good model should have all the needed knowledge or an edit mode
1girling for years to come
>you'd have to fight 25 dollars against ugly bias even harderlol
>>8843672you're in the wrong thread
>>88436721girl is eternal after all
xer age?
>>8843678>>8843675
>>8843677Indeed
hgg2 when?!?
hgg2 with loli
>>8843597Thoughts?
>>8843713We are on cuckchan, there is nothing to do
>>8843713>loli gets banned immediately what caused this?
>>8843720loli matters to advertisers, same with beast/furryboard quality is a second-rate concern
>>8843723grimmest... at least we have cuckbox...
WAKE UP BABE NEW ANIME PREVIEW JUST DROPPEDhttps://huggingface.co/circlestone-labs/Anima/tree/main/split_fileshttps://civitai.com/models/2458426/anima-official?modelVersionId=2764263
>>8843733>still 512x
>>8843733qrd?
>>8843740He hasn't changed the notes yet, so no idea.
>>8843733i'm cumming
>>8843733You're telling me the newest anima partial epoch just dropped? It's really starting to lock in!
>>8843746This but unironically
jordach status?!
>>8843742>The preview2 version is a small upgrade to the first preview.>>A significant part of the training is redone with different hyperparameters and techniques, designed to help make the model more robust to finetuning.>It is trained for much longer at medium resolutions in order to acquire more character knowledge.>A regularization dataset is introduced to improve natural language comprehension and help preserve non-anime knowledge.>It has the same resolution limitations as the first preview. It is trained only briefly at 1024 resolution. Going much beyond this will cause the model to break down.>This is a base model with no aesthetic tuning. It is designed to be wild and creative, with the maximum possible breadth of knowledge. It is not optimized to produce aesthetic or consistent images.
>>8843764Sounds like he addressed the issue with "lora forgetting" (although this really only seemed to happen with certain loras)Will be interesting to see how it handles loras now
>>8843764>It is trained only briefly at 1024 resolution. Going much beyond this will cause the model to break down.WTF? How is he going to go beyond 512p then?
>>8843768>The final version will have been trained on a dedicated high-res phase. Details and overall image composition will improve.
>koku's style is actually slightly recognizable nowwe're so barack
>>8843772how true is this, sirs
>>8843777look at /hdg/
>>8843733feels like an objective upgrade pretty much across every seed ive tried
>>8843782Yep, cant wait to see how it handles lora training. CBA to train tonight though.
what's the best sampler/scheduler combination for anima?
>>8843786Euler a, Normal, 25 steps, 4 CFG just werksTM.
>>8843786DPM++ 2M SDE, 30 steps, SGM uniform. But really you can use er_sde, euler, euler a...
>>8843578Now find the line where assigned to "pre-blured" image variable is being used.
>>8843733>A significant part of the training is redone with different hyperparameters and techniques, designed to help make the model more robust to finetuning.>It is trained for much longer at medium resolutions in order to acquire more character knowledge.>A regularization dataset is introduced to improve natural language comprehension and help preserve non-anime knowledge.>It has the same resolution limitations as the first preview. It is trained only briefly at 1024 resolution. Going much beyond this will cause the model to break down.>This is a base model with no aesthetic tuning. It is designed to be wild and creative, with the maximum possible breadth of knowledge. It is not optimized to produce aesthetic or consistent images.Anyone got any x/ys?
>>8843789>>8843788i see thanks, been using res 2m and beta57 cuz i saw it in a pnginfo that looked decent, but i didn't know if i was just being a schizo because i can't find any distinction compared to default er_sde and simple. maybe im just blind though.
>>8843792of what, nigga, it's a base model
uhh anime preview 2 gens?
>>8843794of anima preview 1 versus 2 obviously
>>8843795mmmmh nyo
>>8843796newer = betterit's obvious
>>8843788>>8843789one of these, anyway. depends on what kind of style you're going foreuler smooths things out and works better for actual anime, while dpm is messier and is normally the default for ptohorealism, also helps 2.5D
okay the small details are for sure improvinglooking forwards to the full bake
>>8843803yeah, soccer is not quite there yet
>>8843805That's kind of insane wtf
>>8843806just write
NAIggers in fucking shambles, this is unironically at least as good as v4.5 already.t. former NAIgger of 2 yearshttps://files.catbox.moe/huciql.png
>8843808lol
>>8843739if it's at 512 how is it ever going to get better, surely all this training is wasted when moving to 1024 or higher
An anima image using ER SDE with 32 steps takes me around 17 seconds to generate, that's not too bad I guessTime to some experiments
>>8843817>surely all this training is wastedIt's the opposite, when you moderately increase the resolution the model quickly adapts to the higher res, preserving everything it learned at low res. But now the model can "see" details only visible at the higher res and starts learning them. There's a reason every single diffusion model trained from scratch does the 256 -> 512 -> 1024 progression.
>>8843820source: it was revealed to me in a dream
>>8843822that was rude
>>8843823only if he's wrong
Is there any significant loss of quality going from 30 steps to 20 steps? I can gen the second significantly faster and I haven't really noticed a big difference
>>8843829I use 25
>>8843808
i_have_no_idea_what_im_doing.png
>>8842227Added a bunch of other things that probably is messing the whole gen but I got pretty closeDid you really struggled to do this, chat? lol
>>8843837Why are you brown?
>>8843839see >>8843772
>>8843839I was meant to become one of God's most strongest warriors
>>8843834I can see melon22, but once again without his amazon bodies
pancho...please...I...
>>8843790good point, seems someone didn't update the code commentary.But hence why I said apparently, I didn't do a thorough check.
good morning sirs
Alright fine, I am just testing this second preview and didn't even touch the first one but, this is so easy, what the fuck Inpaint kind of work too>>8843854Forge neo is not that bad if you are only going to use anima
Ancientautist here.I just want to report that Anima works really well even with lowres screenshots.
wow so much sexiy text and signs! local is saved!!!
So far my biggest issue is the usual Danbooru shit of making the male's skin always significantly darker than the girl. My current theory is the introduction of e621 data is helps avoid this, it's not too bad on noob and iirc wasn't as bad on Pony either.
>>8843883was that an issue on pure illustrious though? I don't remember
>>8843884not sure either, it's been too long and I don't feel like redownloading just for this one thingit is kind of a pain though
Anima wonhttps://files.catbox.moe/1lt3kr.png
>>8843881you mean, as taking an image as a source for a character or something?
>>8843887One of the main issues I had when training loras was when my dataset was fucking tiny in terms of resolution. As in 720 x 480 kind of tiny. I literally have the best resolution available, and I know it's barely acceptable.
>>8843886>wonwhat? there was a retard spamming shit like this genned on illu 0.1 without loras
>>8843891can't you just animesharp it to 2x
>>8843893I did. But the damn thing hallucinated too much because of the blur, and got even worse results.
>literally all the brownoids are here nowwe had a good run chat...
where
>>8843896They're in the Laxhar discord sir
AIDS?
>>8843896daring for a mexican
>>8843877Good morning sir (≧▽≦)
that's a really big turd
>>8843910why Sailor Mars is such a slut for BBC
Seems like the best way to do multiple images/komas/etc.. with Anima is by prompting with the format "1st image: ..." "2nd image: ...."
Is there an easy-script fork that can train on anima?
How do you upscale dis shit
>>8843926nai
>>8843926It breaks down when genning somewhere roughly above 2.3 megapixel. If you need to go higher you can use tiled upscaling, that's ultimate SD upscale or multidiffusion.>>8843927 I think he means noobAI there's been a couple newfags in /e/ lately using the acronym wrong
>>8843929i meant novelai
Remember when the best model we had was pony. Holy shit that was grim.
>>8843931>Remember when the best model we had was ponyI don't.
>>8843929Yeah multidiffusion with mixture of diffusers seems to work okayish as cope
*that was peak
>>8843932so true naisister
>Anima updates>It still can't do jinbaori and aquiline nosePain, I guess my jap and jewish characters will have to wait
>>8843923I wouldn't say it's quite there yet, but it can be done. Reliable though? No.
>>8843931do you member when we had to stack 3 loras on top of a shitmix just to get a "decent" anime style? I do
>>8843935>still seething
>>8843938nah you'd just train the three loras into one
>>8843940That gave birth to many an abomination.
>>8843938>Autismmix
>>8843942reweik actually
I am on a old SD forge build, should I install the new one for Anima?
>>8843925yes, the main onehttps://github.com/67372a/LoRA_Easy_Training_Scripts
>>8843954box please
>>8843946thanks, I guessbroke my install trying to update to this and now neither one works, can't find a torch.dll in the sd-scripts backend venv, even though it's right therefml
But can it do farting on prey?
>>8843963fart has 260 posts so probably
Anima 2 status?
mogs nai
wide shot is wide
>>8843962Turns out you need a very specific version of python, below 3.11 it won't install a bunch of dependencies but above 3.13 it won't install torchThere's a check in the code that's at least supposed to warn you about the minimum version, but it doesn't workit also downloads a version of torch for the unsupported 3.10 so you gotta redo it by hand
>1girl, standing, TEXTThe power...
>>8843990Kino background
>>8843990box
>>8843990is this nai 5?
you localtards cant even imagine how good nai v5 will be
nais v4.5 = pony v7nais v5 = pony v6
any cfg snake oils for anima? the colors and styles burn at cfg 4
cunny oiling my snake
>>8844013You must be doing something wrong. Post an example catbox with metadata.
>>8844013>colors and styles burn at cfg 4it just looks like that, sorry bro
i know this is wrong place to ask, but how do you theoretically solve artist mixing with llm based models without loras
>>8844051that's the neat part, you don't
>>8844052this exact mentality is why local is in the state it's in.
>>8844054nai has 10000 times the compute and an infinitely more knowledgeable team compared to local bakers and they still couldn't figure it out, what makes you think local has a chance?hard to compare with proprietary sfw models like nano banana since those don't really learn artists but i'm personally not that impressed when asking them to use an image as a style reference so idk
>>8844051an architectural layer dedicated purely to style. it will never happen as not even the 100b parameter API models can mix styles properly.
>>8844051>>8844059didn't flux have both t5 and clip encoder? couldn't clip be added for styles specifically? i don't actually know this shit though so i might have said stupidest thing ever
https://civitai.com/models/2438778/anima-real-anime-cat-ara-cat-anima-and-il?modelVersionId=2742074What the fuck is wrong with the people who post and use this?
>>8844059why is that the case? maybe nobody bothered with that? why can't these models, even dit ones, differentiate between styles with text? how about establish a reference (like a realistic image) and make the model learn difference to separate what style is, then just set the descriptors with artist name. that way you can even maybe set specific image attributes to different artists. (like eyes by *artist1*, body by *artist2*.)
we'll still be training loras in 10 years. it is what it is.
>>8844064near instant better ip-adapter when???
>>8844062You don't like 2.5 glossy painfully obvious sloppa, anon?
>>8844067yeah, i'm not brown
>>8844062from my experience with slopper threads/places, it seems like most people have zero knowledge about anime art. I remember some people unironically ask whether danbooru really existed before ai. so i guess for someone whos dick might be attracted to cartoon/anime characters but they don't give a shit about styles this might be the best and easiest choice? i dunno
Whenever someone makes that kind of statement, I can't help but feel that the poster is so melanine rich they shine.
>>8844063>why can't these models, even dit ones, differentiate between styles with textbecause models have been stacking clip-art comprehension since SD3. Z-image turbo:https://files.catbox.moe/gjse2q.pngleft:>a mystical creature which is a fusion between a cat and a birdright:>A mystical hybrid creature that is a fusion of a cat and a bird, sleek feline body covered in soft iridescent feathers, feathered wings emerging from its shoulders, long elegant tail tipped with plumes, sharp yet graceful beak blended with feline facial features, luminous intelligent eyes, perched on a moss-covered branch in an ancient enchanted forest, faint magical aura surrounding it, soft morning mist, cinematic lighting, ultra-detailed, fantasy creature design, highly detailed textures, 4k, concept artnu-models don't understand how to blend concepts, only stack them. you need to 'enhance your prompt' and describe every feature, then watch it layer them like clip-art as you see with the bird beak just pasted on top of the cat. you can try this with anima as well, you'll get the same case of features just being pasted on top of each-other to force rigid comprehension.so just like with concepts, the model just stacks artists which leads to one overpowering the other instead of them blending like with SDXL.
>>8844051It's already solved, you don't need extra steps for this.
>>8844073this is also because turbo is super rlhfed, aesthetics training kills creativity as can be seen by turbo vs base
>>8844074where
L O R AORA
>>8844051Train them for infinite epochs until they learn concept separation
>>8844077eat shit>>8844078hi astralite
>>8844079i can't believe it's not CLIP!
>>8844062i have no right to complain because my current anima style mix looks just the same
>>8844086chatter wtf is that
>>8844086I really like this style.
>>8844076In the model.
>>8844087you dont like horribly deformed jirai kei girls?>>8844088the aco comes for us all eventually
oh damn, anima can even do real life photos (score_9 goes in the negs, fuck ai slop).
I still like animayume more.
>>8844094
>>8844094your animeyume gen sir?
>>8844094you don't need to declare you love eating shit, rajeesh
>>8844096Yes, my gen.
>>8844094hands status?
>>8844098wow such ponyv6 vibe
>>8844098wow is this really ai?
>>8844098In all honestly that body type is hot but the style look way off
>>8844102it's score_9, score_8, score_7, score_6and shitmix, of course
>>8844094same. it's more coherent. tried anima preview2 and it kept giving awful angles, character from behind despite specifying face, etc.
wai-nsfw-anima troglodytes, go to /hdg/
>>8844103>>8844102>>8844101>>8844100>>8844099>>8844097>>8844096>>8844095Yes, and, nogens?
>>8844109
>>8844110Why are you showing me your photo?
>>8844111It's NAI
anima v2 cunny, where is it
nai cunny doko
>>8843818Sounds about right. I think I get it at 13secs on a 4090. Some anon posted some commands here >>8840528 I still need to try.
>>8843829On anima? So looooong digits plan!>anima needs stepsdigits plan!>anima needs stepsdigits plan!>anima needs steps
>>8843854Based holdover. That nigga ain't and/or can't update reforge to support Anima due to all the added bloat/features of reforge. Neo alone freaks the fuck out and nukes itself if you try to add any of the few good things of reforge to it. Too many incompatibilities unless he started from scratch.
>>8843886>grant us eyes, grant us eyes. plant eyes on our brains, to cleanse our beastly idiocy
>>8844079I type like this and post like that on x.com
>>884382910 steps shouldn't make that much of a difference in speed.
>>8844131anima isn't worth the hassle. cumfart is still awful and only ever gets worse
>>8844139The relationship between steps and time taken is completely linear.
Any idea how a 1024 training resolution works when provided with portrait or landscape images? Does it scale to shortside, longside, or to the same number of pixels?I'm getting:>RuntimeError: stack expects each tensor to be equal size, but got [4, 128, 88] at entry 0 and [4, 128, 104] at entry 1Which looks similar to the error you get when genning and your resolution is not a multiple of 8. I went over my dataset and the original sizes are fine, though many are above 2K and probably get scaled down to something bad.
>>8844161Never mind, latest easy-script is fucked. I didn't even have anime selected, when I do it won't load the model.
>>8844062it's for ironic weebs that don't actually like anime, so they need a hint of realism whether that is in shading, lighting, anatomy or facial features
>>8844163Tag me next time you gonna pop off, lil bro.
>>8844086blud what the fuck that's an ayylmao in a skinsuit
>>8844165its a style
>>8844166no that is clearly a foreign creature trying to pass themselves off (poorly) as a person. it doesn't even hit uncanny valley.
>>8844163>ironic weebs that don't actually like animetouhou and vocaloid sisters, our response
>>8844169>2hu is a blacked dogwhistle'fraid so
>>8844169anime is objectively fucking boring, I do like anime art doe.
>"X is objectively Y">look inside>"X is [headcanon]"
>>8844163I don't think it's an inherently ugly style, just really tired of seeing it
>>8844173lol
/hdg/ has been recovering lately
>>8844177ahh ahhh mistress
>>8844173based illiterate retard
>>8844180i'm on 320x240 screen please understand a and e are basically the same
lil bro crashed deleted his tweet
>>8844181it wouldn't be funny if not "we don't like your kind here" and rent free WAI
>>8844175I think it's inherently ugly, looks like those shitty "draw ur animu" tutorial books but with extra realistic shading.
>>8844187>looks like those shitty "draw ur animu" tutorial booksit actually doesn't.What it looks like is airbrushed pinups for chinese bootleg kusoge from the 2010's. Beyond just the aesthetics being absolutely godawful, the reason it incites such a negative response is because it's associated with poor, mass produced low quality ripoffs made purely with the intent of stealing credit card information.
>>8844163You are wrong. It's not about realism, just pure anime and japanese artists suck.
so much this
>>8844165an ayylmao that has boobies!
>>8844191>japanese artists suckthis but unironicallybut for completely different reasons, acojeet
>>8844216> but for completely different reasonsName few, niglet.
>>8844219chinks and gooks are better at anime illustrations now; color use, better proportions, interesting styles without looking aco and composition that isn't just white background or girl being fucked on a bed same with 3d models and character illustrations in games
>i think this style is better!
>>8844223So you are an acojeet, because these are not any different reasons I would name, other than tying to chinks and gooks.
>>8844227lol
Damn it's only 80 seconds per slop on my potato 1660ti. dmd2 is fucking black magic for poor fags like me.
Anyone try training characters yet? I picked an epoch third of the way in and it's more accurate than noob was after the full run.Still can't really control tentacle hair with NL though. Maybe next arch.
>>8844235now stop using slop checkpoints or >/hdg/
>>8844237more like stop posting slop outputsthe end result matters more than the tool used to make it
>>8844237Yeah I know it's slop checkpoint. I'm trying 102d custom but still haven't found good artist mixes yet
naibros we're so back
>>8844240fuck off, pig
creature meltie
>>8844242why are you here
>>8844246Why isn't everyone here? We have a new toy to play with.
ngl the creature getting immediately extremely angry at the mere mention is pretty entertaining. Gonna poke it some more
brown monkey behavior
NAI 4.5 is so good bwos, local could never
>>8844251yeah you're pretty much an angry chimp lol
naisir? it is good, indeed :3
>>8844257it likes the bbc doe :/
Regarding the upscale on anima, plain upscale using something like 4x-AnimeSharp (the regular dogshit version since neoforge doesn't have the implementation for the RCAN version) and then inpaint the bad parts is going to be the right play until the model gets more trainingBut is fine because chat has been using this same exact workflow that I told you since years ago and everyone already have plenty of experience on it, right? :)>https://files.catbox.moe/9ym7i9.png(The gen end up being over 10mb after all the inpaint work wtf)
>>8844280just use mixture of diffusers
>>8844280Seems like I fucked up this one since I upscaled for x4 instead of x2 that's why the file end up being so big oops, but you get the point>can you do [thing] nowNot that great of an example but I also did that on this one >>8844032 (didn't really inpaint more than her face tbqh)Also I think current anima is not doing great at plain solid colour backgrounds but we all know already the hacks for it, right?
>>8844282why are her palm creases noticeably lower resolution
>>8844284Because I inpaint them together rather than one at time, wanted to see how would it go compared to inpainting on SDXL but overall seems like the same Good eye
>>8844280can you post the upscale workflow
>>8844286Keep in mind I am using neoforge, it really only is>good base gen >send gen to the extra tab and pick your prefered ESRGAN, DAT or RCAN upscaler to upscale by 2 >inpaint to fix/add detailsGood ol' "upscale before inpaint" but now we do plain upscale instead of hiresfix
>8844294why is it so retarded?
don't be mean to genuineposters now
>>8844294Anon just use this, it works.
>>8844296you can't stop me from being mean to mexican brownies
noodle schizo is a nai shill, makes sense why he's so braindead
>>8844297awww i'm on comfy
It is worth transferring to anima? or should I wait sars
>>8844302yes
>>8844295I have hundreds upon hundreds gens using that workflow, chatI have also tried CN tile upscale (not available at the moment for anima of course), this >>8844281 and none of them produced significant better resultsI am just sharing experience (again), but if you already have solved that on your own means then I expect way better anima gens from now on :^)
do you guys have some artist you like testing loras onit feels like every one always has some retarded shit screwing over the datasetcensors/low quality images/1 character/no nudes/color bias/ultra aco/low amount of imageslike im currently putting together ririko and this nigga unironically needs to do more 1girl standing
>>8844307dishwasher1910, ratatatat74
>>8844306/h/?
i knew noodleschizo was retarded but i didn't think he was "just use 4x-AnimeSharp and nothing else to upscale" levels of retardedbrownie moment
>test nais>all my artist shitmixes do look like liquid shitThis shit is so bad but on point for some reason
shit, it's only been a day and I'm out of ideas, back to cowgirl pov tier>>8844300Comfy has a multidiffusion custom node in https://github.com/shiimizu/ComfyUI-TiledDiffusionLast I checked it didn't work with anima but that was like a day after release. Ended up using Ultimate SD upscale, which is very similar just older tech.>>8844306It's the same thing in effect. Except automated and across the whole image tile by tile, instead of just the areas you think to run an inpaint over.
is normalized attention guidance going to be compatible with anima or is it obsolete already?
i think i've tried enough anima preview 2 gens to conclude i still don't like how it mixes artists. i'm using euler a + normal + cfg 4, but i tried other combinations (dpm 2pp sde with sgm uniform, euler beta +normal)
I love anima bwos. (But it's slow af)
>>8844350you'll get used to it, just gotta stop expecting it to work like SDXLdiscover new mixes and techniques
really wish we had anything as good as nai. feels bad to be constantly so behind on local
>>88443532 more weeks, trust the plan
5 free gens have been deposited to your account
>>8844350Artist mixing as we know it was just because of CLIP
>>8844342It just seems like a sidegrade to negpip, have you found any usecase for it?
>>8844359it let me isolate clothing elements on characters that didn't have a lot of variation in their artstyle. normally they would always show up with wrist wraps or bandaging, and negative prompting it out required such high strength and artifacting, but with NAG it fully removed it. that's my two cents on it. dunno how it compares to negpip though. i guess i need to just temper my expectations on what's supported by anima since it's so new.
>>8844051Just force tag bleedover whenever the tag starts with @
>>8844367>tag bleedoverQRD? Do you mean prompt scheduling?
>>8844370KekAfaik artist mixing worked because CLIP was so dogshit that tags would bleed into each otherSo, tech illiterate as I am, I thought a solution would be to force this behavior on select tags using the artist syntax from AnimaIf this is even possible I have no idea but on my peanut brain it makes sense, I could see it as an extension
Daily reminder that this thread is for lora and lora related topics.nai >>>/hdg/slopmixes >>>/hdg/pedophilia >>>/hdg/faggotry >>>/hdg/race fetishism (including fantasy races) >>>/hdg/
what character to gen is the biggest shitpost by default, non rule breaking
>>8844380the latest fom gacha whore
>>8844381hmm I think that is too niche and would also have to imply effort via doing it with a lora.
>>8844382fom implies that it isn't niche and that it already has a lora on civitaiare you a gachafag or why are you getting defensive
>>8844383no I am just a high effort shitposter, I want to post the best possible bait image
>>8844384in that case i'd go for blacked or churled yvonne i guesshttps://civitai.com/models/2388381/anima-yvonne-or-arknights-endfield
groidy has fallen
are turbo loras dogshit?
glad I went with NAI
For nai users: what's the difference between those free gens and the actual paid version? even if i input the same prompt + seed + steps + cfg etc. it still doesn't give me the img. Is it nerfed?
>>8844400aura, the image is just worse because the latent space knows you are poor
Conditional prompting is quite good, but I would be lying if I said it is yet satisfactory. Obedience is what I like the most about anima.
>>8844408Is that Roll? I wish I had an oppai loli sexbot.
>>8844411I mean, nobody asks what happens to Roll after Dr. Light dies of old age (and feature creep) in the year 20XX (+Y).
>>8844408it's cool, but I like how it follows styles even moreand it seems to train very easily
https://files.catbox.moe/14fnud.pngPrompt following test
epstein island test :skull:
her age? x3
>>8844447
>>8844448not just that, he made them intentionally much younger
Has anyone trained a lora on the new anima preview yet?
>>8844461no because it won't be worth it
>>8844461Yes, but only because some other anon taught me how. Couldn't figure it out on my own.
>>8844461yes>>8844236>>8844463it's the same as sdxl, you just tick the anima box and provide unet/te/vae separately
>>8844472People kept talking about "knowledge forgetting" with the previous preview. Have you noticed that testing your lora with the new one?
can someone post a GOOD anime gen?
>>8844478>>8844446
>>8844483poor upscaling otherwise pretty good
it's anima
>>8844476haven't noticed that before either, so I might not be the best judge
To the guy with the anima artist preview gallery, thanks. You're a fucking pedo btw
2x-AnimeSharpV4_RCAN is pretty good for img2img upscaler when doing the multidiffusion stuff. Anything else worth trying?
>>8844523it's no different from regular hiresfix, just does the AI pass after upscaling in small tiles instead of all at onceI might be art-blind, but imo as soon as you have 0.2 denoise the differences between upscalers get wiped away anyway
Sars why am i getting double text with gibberish? ;^)
https://youtu.be/IsNXLrTaJ5o?list=RDIsNXLrTaJ5o
poop creatures go to /hdg/
:^(
>>8844515i am a perfectly innocent hag-loverhttps://files.catbox.moe/d1dcwk.png
C U N N YUNNY
is love
Man, I think anima has good styles to be honest but a single artist style is never going to give me the body type and features I want on themThe NL prompting is pretty decent but is failing me at doing super specific positioning of subjects. maybe I am just describing it wrong or maybe it's just not there yet but I am hyped for future improvements
I CAN'T FIND A GOOD ARTIST MXI
>>8844446Based lolification chad
Mixing is a pain in the dick
bibici pls
>>8844583I've found that mixing tags with NL and making prompts VERY LONG can help it with really complicated stuff. Seems like in general the longer the prompt the better.
>>8844590wrong thread, all bibicis in /hdg/
>>8844592ummm anon, don't look 10 posts above that one
>>8844591That's what I have tried, I'm describing my scene as detailed as I can, per character and overal setting, sometimes it work and sometimes doesn'tBut again, maayybe it needs more time cooking before doing those really complex promptsIf it wasn't for the style mixing I would 100% use it as it is on daily basis
What's with Model chefs and their fetish to put weird shit in the preview generations?
>>8844604It's just for "WOAH" factor
>>8844604it's obviously playing on a standard nl test prompt newfriend
>>8844605I was gonna say you are right but I remembered the noobXL preview image
>>8844608the fried Mita? kek
I miss the pony hash mines...
>>8844613In the end those were for nothing but exposing astrakek retarded bake
>>8844614and now that grifter is nowhere to be seen. He used to show up here jerking himself off like he was the messiah.
>>8844614yeah but it was fun and the thread still had a little bit of sovl at that time
>>8844604>one of the key capabilities of my model is natural language comprehension>should I demonstrate that ability in the showcase images?>no, I think I will generic 1girl slop using tags instead
>>8844623>1girl standing:(>1girl standing 4 arms:)
>>8844623To be fair the first image is showing off the NL, the style is just ass.
>>8844623retard
Holy shit anima is comically bad at doing noses>>8844586Same
Prompt scheduling is pretty kino for style combos with Anima desu, just have to do lots of testing to nail down what you want.
proofs?!
>>8844633it was revealed in a dream
cope scheduling
>>8844632Are you on Forge? Because you can't prompt schedule in base comfy.
>>8844637Using Haoming's Forge Neo, works fine there.
>>8844644Why is she orange
>>8844644her cfg?
>>8844632Would be nice if there was a way to control when a prompt enters AND when it leaves in the same gen
>>8844644>>8844645>>8844646>>>/h/hdg
heeeyy guys, anons said this is the nai general, can i post my novelai gens here?
>>8844649Only bibici nai gens are allowed here
>>8844648Answer the question.
>>8844653it's jeet
*ahem*hgg2 when?
>>8844659you're free to create it right now
Don't ever address me with such disrespect.
shitposting with gens is still shitposting
Only nai allowed from this point.
how do i get nai to give me that signature bbc patina?
>>8842660that's cool bwotried it on some artist mixes and found some nice tangential styles
>artist mixesngmi
>>8844668I find it weird how it absolutely cannot tell a 3D render from a drawing
Has any anon here successfully trained a character lora for anima? The model seems to refuse to learn anything for me.https://litter.catbox.moe/71gvdsry9dlhllst.toml
>>8844675Kind of. Poor success. But I assume it is my lack of understanding of the model, since I see others make consistent and reliable models.
>>8844675You can try my settings >>8841486 they aren't perfect but they're passable for a starting point though i havent tested training on preview2 yat
>>8844675I was actually surprised how well it trained the details. https://gelbooru.com/index.php?page=post&s=view&id=4433078Here's my toml https://litter.catbox.moe/ylhf1twhszp9721z.tomlKey differences: you're training the TE too, you're using the warm restarts optimizer but don't actually restart at all; I used LoCon instead of Lora which means training the Conv dims as well. I also did not prune any tags and used an excessively high dim size. Not saying any of those are right/wrong to do, just places to look at.Less important ones: you also got some noise offset, and very different bucketing. I forgot to disable z-snr, although that's probably ignored without v-pred.
Compared to SDXL how much time does it take to train on Anima?
>>8844688Same settings? About an hour when it used to take 40 minutes. Guessing that will depend on how much spare vram you had. As for whether we should be using the same settings, and whether it learns faster; that'll take a while to optimize on our side. Not sure it's even worth trying for a preview bake.
>>8844688take whatever your current speed is and multiply by 1.5xwon't be 100% exact but it'll probably be pretty close.
>>8844688>>8844691*current timeas in it will probably take 50% longer.
>>8844681sheeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeesh
>>8844681can you do one with her giving him a blowjob in the outfit on the leftpreferably irrumatio
>>8844662need to mating press this brat
>>8844701wrong thighhighs color
>>8844644This bitch has wood grain
when will people start posting good anima gens?
gens?
acoma doko
>>8844676>>8844684>>8844685I guess something is wrong with my lora_easy_training_scripts setup because the LoRA isn't affecting the output at all. Tested on both SwarmUI and forge neo.
remember, score_9 goes in negs
score_8 is aco
>>8844730box lorawe can try it out, and the file has full metadata
Good morning bloody benchods :P
>>8844735Here https://litter.catbox.moe/jnggpkvotqf73q0t.safetensorsnikke_mast_tea_service, pink hair, ponytail, bangs, hair intakes, long hair, hair between eyes, pink eyes, heart-shaped pupils, fang, skin fang, huge breasts, cleavage, hair ornament, hair bow, blue bow, hairclip, hair stick, necklace, single earringred kimono, hakama, hakama skirt, beige patterned maekake, hip vent, wide sleeves, long sleeves, shimenawa, geta, side-tie panties, no bra
why does this filthy nigger keep shitting in every thread?
post hints of realism
>>8844751Less ack outs, more gens, sir :3
>>8844756i'm not posting my stuff next to literal feces
pls leave /hgg/
>>8844740I see a tiny difference between 1 and 0 strengthDid you check the loss graph?
>>8844768I didn't enable logging... I will check it tomorrow, thanks
!!! REMINDER FOR LORA TRAINERS !!!
Prodigy it is then
Local is saved>Nekofantasia is the first-ever diffusion model for anime art generation built on Rectified Flow technology, based on the cutting-edge Stable Diffusion 3.5 Medium architecture. https://civitai.com/models/2460560https://huggingface.co/Nekofantasia/Nekofantasia-alpha
>>8844775how do you disable llm adapter training with easy scripts?
>>8844788who?
>>8844789set llm_adapter_lr = 0 in extra args
>>8844792thanks
>>8844788This looks so awful.
>>8844794it's anime
>>8844794>dataset of 4 million high-quality anime artworks>in another first, every single one of them was personally reviewed and hand-picked by the Nekofantasia teamThat's supposed to be an advantage? Just do full danbooru instead of spending 2 years cherrypicking what you think look "good"
>>8844798They had to filter out the pedoshit though because ... they just had to do it okay!
>>8844798touhou tranniesprobably it just means no porn in the dataset
>>8844800>>8844802I remember like 2 years ago there was some push for a "safe" anime dataset that was curated by hand to keep out "problematic" themesIt wouldn't surprise me if this is the result
>>8844788I'm glad anima is just unfiltered danbooru and is already far enough along to blow the fuck out of all these conceited retard projects
>>8844800just search for -loli thoughbeit?
Can we get some gacha slut (male) getting churled in here?
>>8844809it is also not the full danbooru set
>>8844825is it not? I kinda just assumed it was since I was able to generate a lot of pretty obscure styles and characters
>>8844788>no plastic look>tfw want plastic look and anima not having it is one of my beefs with it
>>8844830just train a lora on pony outputs?
>>8844830just use score_9
>>8844788>Due to the extremely aggressive safety filters baked into base SD 3.5, NSFW content generation is currently almost impossible. However, unlike the base model, Nekofantasia 0.1 can already properly render bare breasts — which means that with further training, StabilityAI's censorship (which turned a lot of people off from their most modern model) can likely be fully overcome.cope
>>8844828We don't actually know, afaik. The readme says "several million" danbooru images, I would expect different wording if it was the full set. If this is the full https://huggingface.co/datasets/hipete12/danbooru-2025-10-05 then it should be 19 million
>>8844830the AOM style of 2.5D, or full on 3D render-like, or just VN CG shiny skin? post pic
>>8844836ah yeah I read that and kind of took that they did some kind of light filtering (deduping maybe?) but there's no way it isn't a vast, vast majority since characters with <100 pics and unimpressive post scores work
his age
I think the commander is in his 30s in Exhilium, but I'm not a GFL loremaster.
i have bad news /hgg/...i removed deviantart from the negative prompt.
>>8844836several means 5-9, according to heroes3
>>8844843You fool. You have doomed us all.
>>8844838I'm pretty sure he included parts of gel as well, although they obviously have overlap.
why aren't you posting cunny? >:(
>>8844853mods hate fun and the site is 70% zoomer newfags who think pictures are real
why arent you posting GOOD anima gens?
>anima>goodlol, lmao even
>gens>gooddoubtful
>>8844836There are lots of clothing tags that don't workAnd other minor things
>>8844859have you tried preview 2? I think part of it is just being underbaked, a lot of stuff that prev1 struggled with has *started* to work on prev2 even if not fully or perfectly
loras are a waste of time just prompt better
>>8844862you're a waste of time *vores you*
ponyv7 status?
>>8844860I did, didn't notice any particular differenceI've said it before but despite having 600+ posts 'jinbaori' just doesn't seem to work, even though it worked on illustrious models
>>8844862Sadly, it's impossible to prompt characters the model doesn't know because they're either too new or don't have any art and there's no Edit mode that could be used to provide the model with a one-shot reference.
>>8844868basde
her age???
24
>>8844866ah, iunno then. might be worth bringing it up on hf if theres a handful of concepts that aren't working when they clearly should be
>>8844866Look at the images tagged with that, it might just need more training because its kind of all over the place (unless I'm retarded).
loras are the only way to have true creative agency over your gens. Without self-trained loras you're just a midwit prooompter ai-bro
>>8844873Could be, but it works with very few hiccups on older models so idk
https://files.catbox.moe/31irhb.pngIs cunny juice /h/ or /d/?
>>8844828i think it looks too good by default even without score tags to just be rawjust try the original naiv1 and you will see
>take 2 week goon break>my gen got opnice
>>8844837like this sorta
>>8844878wdym? I thought that anon was implying it was a smaller subset of the full dataset, if you're implying it has stuff outside of it then that's already confirmed in the readme + it feels likely there's some gelbooru in there
>>8844879If only we could get some highlights to spread that feeling around...
>>8844788>training on sd3 when klein existsalso this probably has lower budget than a bluvoll shittune lmao
>>8844883i mean they said they only trained on a third of their 4M dataset and are already running into budget constraints so yeah
>>8844788>looks like shit>can't do NSFW>uses a failed architectureWho asked for this
>>8844881im saying the massive amount of shit images drags the quality down
do we have eta for full anima yet
>>8844886oh yeah that's possible but I'm wondering how they would determine low quality since it doesn't seem to be based on score (obscure stuff working fine)
>>8844884Calling the $500 they likely spent on training so far a budget is a stretch.
>>8844885what do you mean? you DON'T want a worse version of what someone's already doing (chenkin rf) on a bigger and slower model?
>>8844788one of the devs is commenting on the reddit thread. my read of everything is that they're completely delusional, and also jealous of anima since they throw shade at it for no reason
>>8844830score_9 weighted up lil bro
>>8844887we just got the second preview of it so it's not going to be soon
>>8844892if this works you're gonna be my favoritest anon of the semester
>>8844893>second previewwhere
>>8844894i did x/ys a while ago, the ponyslop shiny look gets stronger the higher you go on score tags, the natural language quality tags e.g. masterpiece do not have this issue
>>8843733>>8844895
>>8844897thank
>>8844880just proompt for an artist that looks like thatgive hintobento a try, it's shiny
>>8844891I read their comments too, and yeah they're very defensiveprobably sunk cost fallacy, wasted all that time and energy on something that turned out shit
>>8844868me in the back
>>8844904why aren't you on ozempic?
Anima2 *ALMOST* understands reflections. Kind of.>>8844685This is much better than my own attempts. Even just training for 500 steps produced much better results than what I was attempting.
>>8844900You weren't lying about that guy being a schizo, what the actual fuck is that thread.
>>8844914but not shower heads kekaroo
>>8844916You phyrric win some, and you valiant defeat anothers.
>>8844917That was terrible. Off to sleep I go.
>>8844916this would make a good concept lora actually
Anima is really bad for my sleep schedule
same because it puts me to sleep when i use it
>>8844880yeah I would just train a lora, unless you can find an artist who does that sort of stylekinda looks like what would happen if you slapped a bunch of AOM on top of asurauser
that but nai
>have used nai for the last 2 years every day to fap>removed paypal, now I don't trust it, I never trusted it, I type loli why the fuck would I give you my cc info and full name and address>try local shit but all my gens look like shitPlease help me /hgg/ I have such blue balls, how do I get NAI quality gens in comfy ui, I followed all the docs, stole all the best templates and this is the best I can do, what am I doing wrong?
>>8844935local simply isn't on nai's level, sorry bro
>>8844937Thanks anon, I guess I'll just buy credit card information off the darkweb.
monkaWeird
>>8844935>what am I doing wrong?The first step is telling us what you are doing. That means metadata.
>>8844935>what am I doing wrong?using comfy
just realized that tianliang is still alive
>>8844942I don't even know what you're saying man, I went through the docs in op and copied the last template from this page;https://comfyanonymous.github.io/ComfyUI_examples/2_pass_txt2img/Same image that makes alpine fox girls gives me this slop
>>8844935If this isn't bait, use this - https://github.com/Haoming02/sd-webui-forge-classic/tree/neoAnd than post a gen attempt with metadata on catbox so we can tell you what to change
>>8844935post gens that you used to do on NAI
>>8844947all my gens are gay, I don't even know if I can post them here, mods have mercy>>8844943>>8844946I have a passable understanding of comfyui because I would use it for wan image to video, I can't even get this other bullshit set up I just get mountains of python errors.
Police, I think I'm getting trolled.
ayyy lmaooo
>>8844945if you post an image onto catbox or litterbox, anyone who downloads it can import your workflow from its exif metadata4chan removed the data if you post an image normally
novelai??????
straightest naitroon
>>8844951Oh here you go,https://files.catbox.moe/s0wcp1.png
hgg2 doko?
>>8844955>cardosAnime_v10.safetensorswhat the fuck is that
>>8844935just download Animareadhttps://huggingface.co/circlestone-labs/Anima
>>8844957I tried to run the template 1:1 with the tutorial but I just get this error
>>8844957seems to be from the comfy example workflow>>8844955use noobAI checkpoint for both stepsreduce CFG value to 3-4prompt should be made primarily if not completely out of comma separated danbooru tagsconsider getting a slopped up model that's easier to use for newbies, like wai illustrious
This has to be an elaborate bit
>>8844959Alright I'll try it>>8844961When I did proper booru tagging like I do on nai my gen got sloppier and sloppier, I tried full sentences and at least got some coherent objects, I'll stick to tagging. I also tried illustrious and noobai but I was getting actual garbage.>>8844962I'm just retarded and the fact I haven't had a proper fap in 4 days now is driving me insane, in the words of the japanese; douzo yoroshiku onegaishimasu
>>8844964>pdate] [Auto] 1 new postOk than, you really need to use Anima because it's the only local model with true natural language support.
what is lil bro doing :skull:
>>8844956>>8841728
>>8844968I don't even use it that often I was just getting really really good results on NAI by doing tags for the characters and text blocks for positions and locations, it's the only way I could consistently get...taker pov
>>8844685>you're using the warm restarts optimizer but don't actually restart at allthis doesn't actually matter, restarts are basically placebo for lora training and I think rex only has the version with them so you can just leave it at 1>I forgot to disable z-snr, although that's probably ignored without v-pred.yeah this is irrelevant for rectified flowotherwise, looks good
aaaiiieeee save us noodleman
>>8844974I figured it's still a cosine-looking curve. If you don't restart you spend 80% of the training with tiny LR, no?
>>8844974>placebonyo
>>8844977pretty sure rawr is a logarithmic curve so it only rapidly falls off near the end.
seems more single artist accurate but the composition might be worse? kate is a bit obscure but over 200 so i guess it's fine. most noticeable is in the higus one where it seems to be pulling in a lot more screencap stuff which was kind of the idea. don't see much reason to switch tbqh. i haven't retested quality tags either so i'll probably do that laterno cherrypicking herehttps://litter.catbox.moe/uq4sdrm25l8rztk8.jpghttps://litter.catbox.moe/63rhs134gwk45sha.jpghttps://litter.catbox.moe/jd75xq3oln6ki34h.jpg
oh and artist mixing is COMPLETELY different...lol
>>8840330this didn't change much as far as i can tell sadly>>8844899i like this artist, although anima is still not producing stuff i like. i'm gonna try genning images and progressively adding tags to see if that helps>>8844925i thought about that, i don't have enough images to do it yet though. i'd only want to train it once i've properly inpainted them. but i think that's also a worst case scenario/last resort kind of thing
>>8844988Is that Mion... or Shion? The ultimate cuckqueaning.
Conditional prompting is a mixed bag of dicks.
>>8844896Do you still have those x/ys?
Using artist tags on top of loras with the v1 preview has little to no effect, but with v2 its a night and day difference.
There is no way to properly control the text on anima
>>8845014what do you mean? I can easily prompt text on certain parts of clothes and body writing
loli, size difference, deep penetration, vomiting cum
>>8845014I'm surprised it can even handle a single word, the unet is way too small for it.
>>8845017yeah, now try to put a text on top and bottom of the image that is not attached to anything
>>8845020(Me) okay, turns out the model wasn't okay with me using kanjis but if I use plain english words it kind of work now
>>8845023Inpainting also works with Anima btw, so can try that instead of wrangling it.
>>8845024yes I know but this is more of a simple experiment to learn how to properly control the elements of the gen through NL
>>8845026Ah ok, I noticed that if you add a border to images you can easily place text on different sides of the border.
Bakariso posted a comparison on civit of knowledge loss between preview v1 and v2 when using artist tags with a lora. Huge differencehttps://files.catbox.moe/196jew.png
>>8845034I thought knowledge loss was supposed to be concepts. This is just styles refusing to mix with the style lora.
>>8845034why are you posting on third person, bakaschizo, you don't love me anymore? is that it?
>>8845038HELLOOOOOOOOOOOOOO
wtf is this sd 3.5 anime model meme. are they stupid?
>>8845034you misunderstood bakariso's esl blabber. that's just a comparison between v1 and v2 of his own lora, both trained on the first anima preview>>8845036wrong
>>8845039Don't "HELLOOOOOOOOOOOOOO", you cretin (if you are the real bakaschizo)
novel ai
local will never have artist mixing again lmaonovelai wins forever
good bait but nai also doesn't have artist mixingprepare to loracope and vibe transfer cope forever
but text on a sing
>>8845008https://litter.catbox.moe/wvbu5ygyauu0l4jn.ziptl;drmasterpiece(ye-pop deviantart cosplay:1.2) worst qualitycosplay is optional but it makes it more anime.i also use @hero neisan in negs because it's obese and sepia.
>spend hours making a dataset>spend an hour baking the lora>find out it's fucking shit and needs rebaked>repeat n times local vibe transfer when?
>>8845061he's the hero /hgg/ deserves, but not the one it needs right now
>>8845063ip-adapterit's shit too, but perhaps try running it alongside your bad lora. might give it that final push
(large breasts:5)
>>8845069her age?
flat chest, perky breasts, puffy nipples
>>8845072this but just the flat chest
>>8845073buds are based, tho
(loli:-2)
>>8845074i banish thee to pakistan
>>8845061>cosplay is optional but it makes it more animeYou mean cosplay photo, in the negatives?
>>8845077What did the pakis ever do to you?
>>8845085pakis are the enemies of india (big boobies)
>>8845095entire world should be the enemies of india
>>8844768It was because I used fp16 training which is not supported by Anima.
>>8845104lmao you didn't notice the loss NaN'ing?
>>8845106Yeah silly me tehe
loli x big hilichurl cock
I've never baked a lora in my life
based nai chad
>>8845124me in the back, hgg2 in the front, hgg pov.
>>8845121why is blud crashing out?
hgg2 doko
>>8845129>>>/g/108367053>>>/g/108367058
>oh hey this artist does really nice eyes>complete slop 2.5D body shadingmfw
>>>/trash/81691348
Why can't brown girls be real
>>8845138I have bad or good news regarding that, pequeño hijo de perra
>>8845138just go to osaka and leave the girls out in the sun for a little bit
>>8845137probably a similar thing was going on in /hdg/.
>>8845137I already knew that. It was painfully obvious. I just dismiss any moron larping as a ghetto nigger.
To all my fellow abandoned by pancho niggas, here's a fork for a tagger that won't break your Neo. I should know, my Neo breaks at the slightest breeze.https://github.com/anxety-solo/sd-wd14-tagger
>>8845148Why are not using someting like TagGUI?>https://github.com/jhc13/taggui
>>8845149I have been since moving to Neo. In fact, I've suggested it to others in the past. That said, I still wanted something under one single roof and finally found a fork that just lets me tab over while genning.
>>8845150Oh, alright, cool
>DAMN my lora sucks shit>retrain it 4 times>find out i was testing on preview1 instead of preview2
>>8845186that's why u leave console open baka
>>8845186Just delete preview1 desu
https://civitai.com/models/1615758/tentacle-lactation?modelVersionId=2771130Concept lora bros...
>>8845190yeah i think pv2 is better from artist replication alone
>>8845192based if it works
>>8845192this doesn't work as is? why lora???
>>8845195woah there buddy are you asking civijeets to use the llm TE for anything but describing how butiful sexy and cinematic volumetric lighting 8k HDR octane render the picture is?
>>8845196for a moment i forgot brownies exist, thanks for ruining my day
>>8845192speaking of which, why aren't tentacles /h/
>open hgg>it's just racist spam and a mexican calling others "brown">close hgg
good, don't ever come back
Is that a reply bot?
>>8845136sounds like a nice prompt for adetailer
I didn't even do anything.
>>8845104Wonder if that means my setup is fucked somehow when enabling the lora changed the output, even if just slightly.
yeah you didn't prompt for an artist, but you did prompt for score_9, score_8, score_7, amazing quality fucking masterpiece
It's not pony either.
>>8845195Give it a try before asking? My first question was just how cherrypicked are the example images. Biggest issue on noob was object permanence, the tentacles wrapped around her boobs would stop or start at random instead of a single piece connected to something offscreen. I'm seeing the same on anima2 at least from a few quick tries.
>>8845192>none of the examples even show what the lora is supposed to do>when you can easily get it just by promptingFucking A
>>8845212Works on my machine just with NL and tags
Real talk, why does civitai make NovelAI shills so mad? Is it the community local community and sense of brotherhood that paypigs lack?
finally a good anima gen
>>8845201>a mexican calling others "brown"cathag now posts here? god, we really do need an hgg2.
>>8845217me in the foreground, tryna be slick, tryna low key peek.
C
unny
>>8845236BASED BASED BASED BASED
Mushroom status?
>>8845247Dark Souls bros.. we'll never get KO'd for the first time ever again..
>>8845249mine was a skeletondidn't notice the ramp going up to undead burg, so I kept trying to reach the catacombs
>>8845262damn nai is good
lot of /hdg/ coded posts today
what? only one poop gen and one wai base style gen
>>8845218>>8845247kino gens
so which quality tags/negs for anima if i despise score_9 slop shit
>>8845287Anima barely needs negativesat most "ye-pop, deviantart"it's entirely dependent on what artist you're prompting foras long as you don't put any of the score_ tags in your prompt you wont see the shiny sloppa
>>8845217I should test how many of the more recent AL girls can anima do now
>>8845289Hindenburg isn't that recent, and illustrious can do her as well.
>>8845290Yes I know, I was thinking about girls like Mecklenburg
Good morning naisirs (ง ͠° ͟ل͜ ͡°)ง
>>8845288This is honestly bad advice, I tested both a minimal negative prompt and a schizo one and the latter was miles ahead every time.
anima bibicipred fucking when?!
>>8845294no it wasn't
>>8845295What's wrong with prompting for beebeesea the old-fashoned way?
>>8845269i can't help but think having a girl sit on your pelvis like that would hurt like hell
>>8845148This actually workedI wonder if there is a model for these that does natural language, that would be fantastic
>>8845345there is, it's called joycaption
>>8845346That's limited and gay
>>8845348true
>>8845345you can actually get pretty decent results from feeding taglists to gemini flash
How are you guys preventing the prompt bleeding between different subjects on anima?
>>8845358I reroll.
>>8845358The more verbose and fastidious you are at describing characters, the less likely the bleed is, at least from my experience.
>>8845358combine NL and tags lil bro
>>8845358specifically use relative direction e.g. the girl on the left is blah blah blah
>>8845358if they're tags the model doesn't want to cooperate with then you have to cope and inpaint
No one has ever done a gladiatorial sex battle comic, have they?
>>8845365i've seen several
>>8845365Not ones that take place in an arena covered in dough at least
>>8845368it looks like a concert hall, not an arena
I'm honestly still genning on 102dI can't let go anons...
>>8845371this but autismmix, it just works
>>8845372this but aom3, the style hits just right
reminder that there literally isn't a difference between 102d and wai
>>8845357nice
>>8845371it's worth the trouble artist mixing to switch
>>8845360>>8845362>>8845361I am doing exactly that, but I'm not sure how to indicate where one part of the prompt ends and another begins, so far I have tried mixing tags with NL, using commas and/or dots in between, but it just wasn't workingNow, what is working better is separating each subject, description, and action with a comma at the very endexample:>1girl, standing, medium breasts, uniform, skirt lift, the girl is leaning a little to the side smiling with a smug face to the viewerwhat I am doing now>1girl standing medium breasts uniform skirt lift the girl is leaning a little to the side smiling with a smug face to the viewer COMMA [NEXT PART OF THE PROMPT]But doing that is fucking atrocious to read
>>8845377really not that bad tbqh
>>8845377I found that doing a more or less complete NL description of the scene (always specifying which character is doing what and where they are) and then adding tags as needed works best. Of course the NL description length depends on how much you are trying to work outside the tag system.
>>8845379It is>https://files.catbox.moe/t3uhv6.jpg
>>8845382seriously i don't see what's so bad about it? just don't type the fucking comma
>>8845351Wdym?
All roads lead to controlnet and regional prompter
>>8845383>______>
>>8845386there's nothing wrong with that
>retard who doesn't know english or how to write nl prompts was noodleschizowhat a surprise!
also apropos of nothing ER SDE is a nice sampler. stochastic bros stay winning
>>8845390Even gringoes don't know how to express themselves properly, It's not my fault None of the people I know or have worked with ""professionally"" gave it any importance
just buy a pen tablet and learn how to vaguely draw already to more properly direct thingsalso learn how to do proper post process and fix shit by hand
>learn to drawI wouldn't be bothering with AI if I was content with that option.
>>8845397Already know and have done that plenty of times, I am testing and playing with the newestestest model and architecture just to be hit with "just inpaint bro" again, not cool
>>8845400until we get a neural link that you can straight up just imagine what you want it to draw and it happens(provided you can rotate the apple - you can rotate the apple, right?) there will never be a model that doesn't end up slapping you with "just inpaint, bro."
>>8845397subhumans who suggest this shit in ai thread should get their teeth knocked outgo inpaint on based65 faggot wannabe artist
>8845407why is it so mad?
learn to use AI legacy artcel
>>8845407NTA, but even something serviceable helps.
>>8845384google has the best vision models, so you xan feed image+taglists to gemini flash and ask it to output a natural language caption that incorporates as many of the tags as possible. if there aren't already comfy flows to batch the process it should be pretty easy to ask whatever llm to write a script to do it.
Is there any local model/LoRA that lets you emulate the wide range of color/lighting NAIv3 had, like picrel for example?
>>8845464anima
>>8845464literally not happening unless they decide to release naiv3's weights (so never)
>>8845464just proompt?
>>8845464anima is actually responsive to prompting light levels and color range but I haven't tried to see if you can get naiv3 chiaroscuro effects
>3 months since last wai versionit's unironically kino I hope it keeps getting better
Don't think I have ever worked with lighting like this before. I *think* it can be done, but I have no idea which particular keywords the AI will respond to correctly.
>>8845488spot colormonochrome etcez
>>8845488Use natural language along with tags, you'll be surprised how creative the model can be with lighting (i.e. you can prompt for red lighting coming from the top of an image, harsh contrast, etc, etc..."
>>8845489>spot colorI completely forgot about that term. Thanks.
>>8845382you are also not using the suggested anima prompt ordering just sayin you may wanna prepend that to your prompt
>>8845501nta but what's the suggested prompt order? quality - artist - character - background stuff?
I think anima fucking SUCKS
>>8845504just read https://huggingface.co/circlestone-labs/Animain my testing it is fairly important (you can omit ratings tags, it gets those from context). and just today i started just weighting artists up to 1.2 and skipping the masterpiece and i like it
>>8845448Using all danbooru tags as a taglist?That's quite ingenuous, I don't really bake loras so I don't even need the script, I just like using tagger to help me prompt for things
>>8845505True, why bother when we have WAI and CatTower sar
what are your favorite interrogator models anons?ML-Danbooru CaFormer dec-5-97527 seems to be fantastic
>>8845511australia is giga cucked
>>8845510>ML-Danbooru CaFormer dec-5-97527Is that so. I thought smiling wolf's slop was the go to? MOAT, VIT, and EVA. Guess I'll have to try it out.
i love this model lmao
>>8845526weird kink, don't stop tho
>>8845527thanks king
>>8844830ai generated is a tag you know
>>8845538Now that you mention it, someone just made a post about that on the Anima HF - https://huggingface.co/circlestone-labs/Anima/discussions/83I added all of these to my negative and it seems to have a noticeable effect on making gens better. I'm guessing that since the dataset is so recent, the few thousand images on dan with those tags weren't excluded.
>The Pony-derived quality side has a separate issue. Pony V6 explicitly describes its training set as being aesthetically ranked based on the author's personal preferences. [...] The practical result is that using score_9 in prompts pushes Anima noticeably toward a Pony v6-era look, simply because those images scored highly on that classifier and entered the training data at elevated rates.>The root cause here is a variant of dataset contamination related to the model collapse problem. Old taggers and aesthetic scorers assign high-quality labels to AI-generated images that visually resemble their preferred distributions. Those labels and artifacts then get learned by the next model as actual markers of quality, slowly narrowing the style distribution.pony is still fucking local over, 2.5 years later
autism-animamix when?
>>8845544The whole idea of quality tags is flawed. Training will always average a single style bias out of it, rather than generic guidelines that could be equally applied to a graphite drawing or a 3D render.
Anima NEEDS style clusters
>>8845550how would that work?
>>8845511I feel like more people need to prepare for the probable cumageddon that seems inevitable at this point, but of course Onlyfan whores won't be effected.
>>8845544
>>8845551like shit>>8845545you already have anima cattower and other 20 shitmixes using ai-slop dataset
>>8845474>>8845466How good is it for use overall/for making LoRAs? I haven't used local models/done any training since Based64.
>>8845558Based64 is like using stone tools, Anima is like having the ability to colonize any planet at the speed of light.
wai anima will save local
wai anima is literally nai v5
why anima?
>>8845542>>8845544this fucking GPO CES loving ponyfaggot ruining local with his ugly fucking ai-gen biased score systemFUCK
>>8845567Nah >>8845554
Oh yeah speaking of old names whatever the fuck happened with Waifu Diffusion did that shit ever go anywhere?
>>8845568let's put nine paragraphs in negative prompt to make the model actually work
>>8845570they made a half hearted try after sd1.5 but got bootyblasted by sdxl and haven't seriously produced anything since, iirc. maybe somebody from their team is still doing stuff idk i'm no discordtranny
>>8845570as if there was ever any reason to think that would amount to anything loleven fucking resonance had better chances than those retards
>>8845575What happened to Jordach
>>8845580had tree fake xir death after spending all of the cascade donation model on a fursuit, hasn't been seen since
>>8845582Based
good anima gens doko
>>8845557>you already have anima cattower and other 20 shitmixes using ai-slop datasetGIVE ME MOAR. I NEED MOAR, SNAKE.
>>8845585good anima, whatever happened there?
at least we are free from shitty vidgens
sorry i put score_9, ai-generated, ai-assisted in negs and you guys wouldn't like that so i won't post it
We only use WAI in this house chud
>sorry i [HEADCANON]
quite a lot of /hdg/ shitposting in this /hgg/ thread...
bake /hgg2/ already >:(
>>8845593I don't see any...
>>8845593>still calls it /hgg/ post #10kekerina of time
pre 10 posting uoh
just write
yellow and orange are my favorite skin colors!
Will anima have functioning artist mixing in the full release? I can't find myself ever using it if it's always this shit, nobody ever just prompts a single artist when genning
it won'tand any new model won'tdeal with it
>>8844371>>8845030>>8845262boxes please?
fuck you
>subhuman /hdg/ refugees fail to assimilate with the culture of the thread and turn it to shit just like their original habitatcrazy how browns manage to do this over and over again in real life and online spaces.
why is it melting
brown poop gens when?
novel ai!!!!
Only catfag can save what's left of hgg
bobblehead toddler anima lora when
>>8845703But /hgg/ culture boils down to refusing to post images while acting like armchair data scientists who know the perfect formula to train the perfect anime model.
>>8845728it still attracted all of you here, so what's your point?
retards really thought going into another toilet would mean there wouldnt be any shitthread splits are so retarded kek
>split thread works for more than a year>thread splitting doesn't workwhat did he mean by this?
>>8845727my pov
>>8845703/hgg/ is brown avatarfag general
>>8845736this. it's time for another split
>split thread works for more than a yearlol
>works>but except when it doesntthen it didnt work retard lmao
>>8845736yeah, just forget about NAI shitposters including the one spamming with turd photos
yeah I think it's time we all move to /slop/
Good morning sirs! let me remind you that those thread splits were the biggest shitposts you guys pulled off on this board ( ͡~ ͜ʖ ͡°)
you know, I kind of miss the inclusion e621it had a lot of useful tags for body shapes, like breast shape and how big thighs you wantwith just danbooru you're kind of at the mercy of the retarded taggers that don't know what constitutes as small, medium, large or huge breastsI also preferred its "nipple outline" tag over "covered nipples" as covered nipples more often gave you some retarded skintight fabric around the nipples instead of just like a slight nipple bulge
>>8845804>I also preferred its "nipple outline" tag over "covered nipples"mein nigga
>>8845804I miss "erect nipples"
retards using e621 tags and lamenting their loss really explains why threads have been so infested with aco shit
>>8845804e621 is absolutely 100% beneficial for nsfw and people who deny this are genning 1girl standing with white background. Whether it degrades aesthetics is still debatable but as we've seen with noob it's not a big issue.
>>8845817I'd agree if anima wasn't similarly muddled with trash data that isn't even useful for degen concepts this time.
>>8845818noob isn't necessarily the best example but naiv4.5 proves that e621 isn't a problem if you don't have a massive skill issue and don't suck at curating (difficulty for local bakers:impossible)
>>8845822You'd have a point if 4.5 wasn't ugly as sin
>>8845822nai might be the best example against e621 though lol. v3 furry was incredibly ugly and 4.5 is clearly a downgrade from v3 aesthetically
>>8845822>naiv4.5>quite literally the ugliest base model available>a good example of e621 not degrading the modellmao
>>8845825>>8845827>>8845828full cope right on cue
>>8845804For me it's monstergirl shit, tonguejob, tailjob, tentaclejob; more accurate tail placement for Tomimi, better hooves for Soraka, etc.more detailed pussy anatomypectoral_bulge, as a stomach_bulge that goes up between her boobswhite_penis, as a hack against brown menand a couple artists
>I'm right because I am and don't concern myself with the majority opinion
the majority opinion is that naiv4.5 is the best nsfw model available lol
>>8845822Even on anima, artist reproduction is amazing despite the deviantart and ye-pop datasets. Seems the model can keep them separate with a simple negative, and often eve that is unnecessary.
nai shill really doesn't like how badly nai already gets mogged by a scuffed 512 res preview bake kek
>>8845833source: trust me bro
localkeks: I CANT MIX
>>8845837That seems to be a problem they both share>>8845835honestly the best part right now is lora training, which you can do at 1024 and already get the benefits of halved VAE compressionNL is nice but I'm used to CN and regional prompting anyway
>>8845843>NL is nice but I'm used to CN and regional prompting anywayterminal localtardation
What board do I post guro monster girl femdom rape? Seems incredibly niche of a fetish since I can barely even non-AI originals of it so not sure what would fit.
>>8845819anima looks better without any artist tags. when you apply an artist, it gets blurry on top of generally looking bad and failing to properly mimic out their style. you can test it out, just prompt with artist tags, same seed, no artist. no artist gen looks better. there's something very wrong with the way anima applies artist/style to images imo
>>8845848it's probably just the 512x compression1024x will fix, trust the process
clip won
>>8845849we'll see. i hope you're right but i'm not optimistic
I want my model to be conceptually pure so I can shit on people who like furry or western stuff
>>8845849just 2 more partial epochs
loras won
>>8845852You'll love the model trained on a million deviantart and midjourneyslop images then
fishnets are still dogshit
>>8845857entirely depends on your sampler and how fine you want the pattern to be
>>8845857How do you even prompt them? Every time I do I get opaque pantyhose and thighhighs with a net texture, not an actual fishnet with holes.
>>8845857it'll be fixed with the 1024x training
>>8845849i think you can't fix it. the artist seems to be just a filter applied on top of an existing image, i don't think it's just the resolution
>>8845847depends on the level of guro, full-on is only allowed on /trash/ afaik but more reasonable ones should be fine on /d/
>>8845848take your meds
>>8845848Are you guys not putting @ in front of the artist tags or something? Anima isn't quite as good at artists as base Noob, but it's close, and way better than something like WAI which destroys styles.
>>8845873he's probably just retarded
>>8845859Have you tried to specify them with NL? I personally haven't tried as the idea just crossed my mind but maybe it's worth a shot>>8845860This is slowly turning into a meme but I do believe it, preview2 is almost there
anima defense force always on call
ugh a brown person just replied to me
>anima is better at replicating artists than a model shitmixed with a million 2.5d shiny lorasthey're not sending their best
>>8845870It's girls being taken over by eldritch abomination parasites that split their heads a la "The Thing". Blood, exposed brain, eyeball, mutated face, etc. I guess kinda like koiso usu monster girls but with more actual gore and not just body horror.
>>8845886probably /b/
>>8845888/trash//b/ doesn't tolerate guro or furries, only loli
>>8845889/trash/ wouldn't want this as well desu, maybe catboxed with a caution>only lolilast time I visited the degen threads it was full of random shit
>>8845891I'm talking about rules, not anons' tolerances
>>8845892>rulesI'm pretty sure /b/ only disallows actually illegal content, it has the most lax rules on the website. I 100% seen furries there
>>8845886b-box?
>>8845894free bmj
can't wait for the new highlights!
>>8845896Sure, beware though:<https://litter.catbox.moe/msgg6ydyqwb905kn.png>
>>8845904I kinda imagine it being more gruesome from your description
Preview 2 is terribly sensitive to finetuning 800 steps at 1e-6 already burned it
>>8845786Yukino is such an underrated BBC goddess.
>>8845907I mean, it's very gacha depending on artist or seed how much is monster vs girl features. Could easily add more gore tags but I don't want to overload the prompt too much.<https://litter.catbox.moe/bxfwnsxwdl1zh9i8.png><https://litter.catbox.moe/tuw5ecn5o1b83urx.png>
>the night realm is actually a physical place under natlankek thats natlanyahu for you
???
uhhh, lil blud?
>poltards crossposting here/hgg2/ when?
>>8845918make a discord group, unironicallyyour "golden era" /hgg/ were just three avatarfags
ded
>>8845913Based yukino enjoyer
>>8845913never heard of her, so perhaps she is rated correctly
>>8845955fr fr skrrrrr. bitch never had a fortnite cameo ong
>>8845960>ongoh Nai god?
>>8845804bokobros have suffered the most with this exclusion.
>>8845973i wish highlighterbro could see this
>>8845974>beautiful background>melted details>cowgirl povchecks out
>>8845975were too lazy to face paint, comfyui crashes my firefox if i use the mask tool so i have to load a different browser
Speaking of inpaint, one thing I don't miss from the pancho exodus to Neo, is that inpainting is a breeze out of the box. Shit would lag and slow down like a bitch on reforge whenever I was masking areas.
Size difference stuff is way easier on animahanging_legs works reliablyminigirl actually makes hand-sized pixieseven lifting_person has more poses
finally the perfect ingredients for more pedo gens
>>8846002could always do huge amazon women insteadleft one in the other thread
>>8845848weight the artist tag up to 1.2 and drop all positive quality tagsno i am not joking
Anyone has had this issue with Anima where prompts using a lora has downright bad hands? I keep getting chaos spawn-tier handblobs. I cannot figure out what I'm doing wrong with my loras.
>>8846016Nope. General advice is don't train the TE, and use about half the LR you would for SDXL. Check an earlier epoch, maybe you've burnt it already.
>>8846016lower the lr but yeah anima loras burn very easily from my experience
>>8846017i haven't had any of these issues but i am also zeroing out all the relevant down blocks because it works better that way for concepts
>>8846008Prompt weighting works really well on Anima outside of artist tags as well.
>>8846033positive weighting works great but negative weighting is a total crapshoot
err fractional weighting i mean
horse pussy x3>>8846036You mean going below 1.0 strength? It does feel a little off compared to sdxl, but I haven't had any major issues, not with styles and characters at least. For some concepts it just goes binary yes/no instead of reducing the amount of "thing", like depth_of_field. You can work around that with prompt scheduling.
>>8846042yea below 1
>>8846042That these 3 Nearl family women are all single takes my entire supply of suspension of disbelief.
>>8846050That's gacha worldbuilding for you. Having one out of 300+ operators be married was already a huge surprise.
>>8845955>never heard of yukinohow old are you? I find it odd someone would find themselves on the hentai board of this boomer website without at least being familiar with her design
oh right, anima can do most of the zzz sluts time to goon
chat, what did he mean?
>>8846060die
why lil unc is seething
Can Civit please just go under already, these sars are already uploading Anima loras trained on pony/WAI outputs.
>>8845835> how badly nai already gets mogged by a scuffed 512 res preview bake kekif by mog you mean broken anatomy awful backgrounds and colors then yes definitely mogged
naiggers hurt by that one
i paid at least 25$ for it, so i need to defend it, otherwise i would look like a fool
>>8845893goodwhy one image instead of splitting
>>8699581
if anima is so good why all the images posted are shit?
>>8846143True, where are the WAI and cat tower gens?
in /hdg/
/hdg2/ doko
just wait for the 1024x bake
these localkeks arent ready for nai v5
i got my 25$ and good boy hat on already!
nai v5 will change ai gen forever
>>8846156GIWTWM
any cumfart users? how do you prompt schedule with anima? comfyui nodes for that suck
They laughed at us having to wait to play with new toy but whose laughing now Neobros.
we are laughing t. nai chads
any nai gens to put localkeks in their place?
>>8846193>>8845786
>>8845786Who split into who?
can we get a split thread just for anima now? getting really sick of sharing the thread with SDXL retards
this but unironically
Based. Pony, Illu, Reweik, Noob, and Naiggers need to leave us Anima chads be in /hgg3/
>>8845034Why does their @as109 output look so different/newer from mine? I can only get a soulful severely-outdated version of his style, even if I use "masterpiece, newest, recent".
>>8846200are you using their lora
@asura /(asurauser/)
>>8846055wizard. there's way too much anime to see it all. plus a lot of it is highschool age drama that you kinda just grow out of
>>8846202this is a really nice artist neg if you don't like outlines
>>8846202>>8846210/hdg/ was a healthy thread when asuranon was around.
>>8846211that's nice, dear
>>8846201No, but those four artists aren't in their lora's style list, so those four styles should be coming from Anima. Yet I've never managed to get anything with @as109 that looks remotely like the lower-left.
thoughts on this snakeoilhttps://civitai.com/models/2435207/anima-colorfix
>>8846216See his noob v-pred colorfix for explanation.
>>8846214idk their metadata looks fine i guess except i don't think positive qualityslop has a lot of influence, negs work way better if you want to control outputs.
>>8846220this style really looks fried on every model, same with m-da s-tarou
>>8846222yeah things that are really strong stylistically tend to do that
>>8846214>>8846220It looks like using both pos and neg prompts to strengthen newer/prettier and weaken older/uglier does move the needle. I thought I tried this already without a lot of success, but that was probably with preview 1. Sweet.
>>8846188Not sure if the Prompt Control repo was updated with anima support already. That one's "PC: Schedule Prompt" node works like on a11/forge, with [a:b:ratio] and [a|b:ratio] syntax right in the prompt. Also any kind of scheduling you can hack together using refiner flow, refine with the same model only with another prompt. Same can be done for loras, cfg or anything else you think of. https://litter.catbox.moe/pachbcy70srg8ieh.png
>>8846239Bit of a messy eater, innit?
>>8846229>It looks like using both pos and neg prompts to strengthen newer/prettier and weaken older/uglier does move the needle.Any specific pointers with this? I also have some artists who changed their styles over time and look far better in recent years but I'm having trouble getting rid of traces of the old style even though I use "newest" in the positive prompt.
Full model when
>>8846239>Not sure if the Prompt Control repo was updated with anima support alreadyhttps://github.com/asagi4/comfyui-prompt-control/issues/137
>>8846244cool, thanks
>>8846242Actually just adding "newest, recent" to pos and "mid, early, old" to neg seems to do the trick okay-ish. Adding them to my schizo style preset.
>>8846242"newest" is just a generic style tag averaged out from all recent danbooru pics. Not the newest version of whatever artist you prompt for.
>>8846250Obviously, but at least in theory, a combination of "newest" and the desired artist *should* nudge the model towards the artist's recent style. It's basically the difference between>give me an average of all the artist's imagesand>give me an average of all the artist's images AND all images from recent yearsThe latter would clearly be more biased towards the artist's recent style as his old images would be weighted less overall.
>>8846252Even if that doesn't work, for many artists the generic "newest" bias is exactly how their style has changed over time. Shinier and with more contrast.
>>8846252it does do thatproblem is newest is muddied up because it's standalone and not structured like @artist (newest)
>>8846255That's really not a trivial thing to solve at all and won't be until the booru starts tagging images from the same artist that have different styles with artist_(new) and artist_(old), which will never happen since choosing where the line is drawn can be tricky if the style evolved over time rather than with one massive shift after a hiatus or reevaluation of identity or some other type of melty. The best you can do with the tools you're given is shrink the solution space such that it includes his new style but excludes his old one. "newest" (and possibly "recent') in pos is one half of the equation, "early, old" (and possibly "mid") in neg is the other.
>>8846257nah just train a lora
>>8846258Fuck off.
>hating loras in the HOME of the lora; /hgg/why lil blud like dis
>>8846260nta. fuck off
Man, how am i supposed to gen "vaginal tugging, anal tugging" without E621 in dataset?
you load up pony v6 and fuck off to /hdg/
this is /hdg2/
we need /hgg2/
loratards are perpetually mentally stuck in february 2023they want to keep training forever even if there's nothing to train and there's no point to it
>>8846262My otoko_no_ko \(forced to use trap on anima\) gens no longer hit the same without that extra friction, chat.
>>8846267They also forget that if there isn't enough reference material, like if it's an obscure character with no art besides the one official piece, you can't even train a LoRA in the first place.
just inpaint
there's literally nothing you can't do between loras, controlnet and inpainting if you're not brown
just pick up a pencil, actually
>>8846269You can, it's just an inordinate amount of effort.
are /edg/ posters sentient ensouled human beings or is it just bots posting random civitai galleries and making retarded posts
>>8846299they're all highlighter tulpas
>>8846307>>8846307>>8846307
>>8846299both
page 11 BUMP
