Previous: >>110063934New anime model edition!https://huggingface.co/Comfy-Org/MOSS-Image/tree/main/diffusion_models>UIsAniStudio: https://github.com/FizzleDorf/AniStudioComfyUI: https://github.com/comfyanonymous/ComfyUISwarmUI: https://github.com/mcmonkeyprojects/SwarmUIKritaAI: https://github.com/Acly/krita-ai-diffusionForge Classic/Neo: https://github.com/Haoming02/sd-webui-forge-classic>Image ModelsIL: https://rentry.org/illustrious_checkpointsAnima: https://huggingface.co/circlestone-labs/AnimaKrea: https://huggingface.co/krea/Krea-2-TurboNovelAI (paid): https://novelai.net>Video ModelsH3: https://huggingface.co/Comfy-Org/MiniMax-H3Turbo lora: included above>How to generate anime:image : https://rentry.org/comfyui_guide_1girlvideo: https://rentry.org/wan22ldgguide (No H3 guide yet)>Style ExplorersAnima: https://anima.mooshieblob.com/ | https://animadex.net/Illustrious: https://tagexplorer.github.io>Checkpoints, LoRAs, Upscalers, & Workflowshttps://civitai.comhttps://civitaiarchive.comhttps://tensor.arthttps://openmodeldb.infohttps://www.seaart.aihttps://www.liblib.art/>Resources:Share Metadata: https://litterbox.catbox.moeImg2Prompt: https://huggingface.co/spaces/fancyfeast/joy-caption-beta-oneOnline metadata viewer SD/NovelAI: https://spell.novelai.devCatbox/Metadata Userscript: https://gist.github.com/catboxanon/ca46eb79ce55e3216aecab49d5c7a3fbOutput cleanup: https://rentry.org/RemovingDiffusionGunk , https://www.mediafire.com/file/vipr23exc5htmnt (batch processing python script)Inpainting Guide from an Anon: https://files.catbox.moe/fbzsxb.jpgcollage >>106520607>/adt/ logo:https://files.catbox.moe/pakfgt.pngcollage
gm saars
First for TAD: Total Anima Death
>>110086899it sucks but i honestly can't go back to illustrious after having even just a taste of a text encoder.
Is there any new model on the horizon?
>>110086971krea3, h3 image, flux3 video, ideogram 4
>>110086993Krea3 looks alright, not as much of an improvement as I was hoping, but still something. Sucks that Krea is more resource-intensive than Anima, I doubt my 32GB of RAM would be enough...
>>110087136surprisingly krea2 turbo takes about the same time as an anima image, though to be fair im using 8 steps for krea2 turbo and 30 for animamoss on the other hand takes about 2.5x as long since the recommended steps is 50
>>110086926What's going on with Anima?
>>110086993>krea3, h3 image, flux3 video, ideogram 4,Oh look, Oekaki found Jesus and his name is Catjak
>>110087178I find the outputs are just not as aesthetically pleasing, on average, as illustrious ones.you have way more instant control over the outputs which I really enjoy, but having it soft gimped as a tradeoff is disappointing.the text encoder is also under-powered. i dunno why they didn't train on a stronger one, even just a 4b. loading and unloading models doesn't eat that much time at all, and the buff to prompt comprehension would be worth it.overall its a cope sized model with a dope setup. the most dissapointing part of it is the untapped potential. this thing feels like it was made to be, and it really is, the best anime model to run on 8 gb of vram.If we could get better gradation, even just a model that can be the best for 12-16 gb vram, the results would be insane but that never seems to be a thing that will exist.
>>110087311Is that an Anima gen?
>>110087317Yeah.
>>110087178Falseflag local model. It's been almost 10 months since the first preview and nobody could make it better(not saying it's perfect though, it's not). It still has gaps, bad styles, no tag weighting and the same bad hands/eyes/lips/feet like SDXL. The dev barely talks to people, super quiet and lazy about it, we haven't heard from him in a long time, for example.
Anima has a good dataset, but it doesn't even compare to the illegally scraped dataset from chinese noob.
>>110087136This is like a painterly Nano Banana, the objects doesn't mesh together.
>>110087353>no tag weightingDisregarding everything in your post because clearly you do not know how to prompt
>>110087353So what do I use if I want to gen copyrighted characters?
>>110087445depends on how niche they are. if theyre obscure but still have like 200 posts on danbooru just use anima and ignore the shitposter. if theyre more popular you can still use anima but could also try out krea2 or moss
>>110087353>we haven't heard from him in a long time, for example.well yeah he doesnt post here he posted in ldg
>>110087466>krea2 or mossdo either of these do artist styles?
>>110087634krea2 does more general and series styles. for moss i havent really noticed if it does. the prompting guide says to use @ like anima but so far i havent noticed much with the artists ive tried
>>110087541Huggingface, civitai, he mute unless aesthetic or turbo slip tune
>>110087668i see it's from twitter, any guess what model made this?
>Previous: >>110063934 (Cross-thread)>329 / 86we need to do better
>>110087727nope. and the genner probably won't tell because they sell it
>>110087668I know what happened with your image. The feet came out mirrored, and you couldn't fix it with inpainting because the model can't read the edge of the image so you ended up with the legs swapped instead.
>>110087735The fuck? He couldn't even be bothered to fix basic clothing details and he tries to peddle this trash?
>>110087668Oh preteen upperbody with hag legs and feet, nice frankestein
>>110087733Tell that to the finetuners first
Lmao tried MOSS, it's garbage. Anons hyping up whatever random local model. Krea 2 for macro and Anima for micro are still undisputed kings.
>>110087667thanks for the reply, unfortunately for my purposes it's no good if it can't do niche styles or self-trained loras. guess I'll skip
Moss imaeg tiem!
More moss imagee
shun secks
Moss imaeg Gemma-chan(reminds me of SD 1.5-type stuff)
>>110088345gemma a cute
Last one
erotic ba girls sugoi
>>110088362I like that it's kinda retarded. It's undertrained with nudity but I'm going to see if my danbooru 40k dataset helps it.
>>110088392it's not *that* undertrained, i've seen plenty of penetration and body fluid gens out there
>>110088345i expected more out of this. the bed looks bad, the furniture beside her doesn't look proper either
>>110088471It's the base model (bf16). It also seems very seed-dependent, which is cool for variety, but you get lots of dud gens.
who tow sex. furiner sex
I'm back on these threads y'all
>>110088379lovely teabags
imagine moss with booru data finetuneit already has some knowledge tag-wise
>>110089332>eyebrows
>>110089831What a cancerous series that was. Haremshit usually is, but that one didn't even pretend to have a reason for other girls to even be there.
>>110089879goutoubun no hanayome was a good series.
Can moss do artist styles?
Wait nvm you niggers said its undertrained on nudity and no e621 training so its DoA
MOSS consistency is worse than Anima, Flux 3, SDXL, it can't even render a straight spear, and sword poses in 1girl images come out completely hallucinated. Krea 2 is still the better model, don't get baited into thinking otherwise just because MOSS happens to know Miku, Genshin, and a handful of 2hus. Sure, it's picked up some Danbooru knowledge, but that reads less like quality and more like a calculated move to generate engagement for a mediocre model. They timed the release to ride the backlash from Flux 3 flopping.
>>110090592Literally who cares about new models until they reverse engineer Novelai v5
My flat wife
>>110090663Fine tuning DOA models on popular gacha game characters can be a useful tool for marketing but it should not be the only focus to say if a model is good or not. Anons are too quick to jump to conclusions here.
>>110090771Until a model can make Namine getting a horse bukkake while Sora sucks on her tits its DOA
>>110090837Yeah, we're in 2027 almost, we can't settle for less than 12b models.
>>110087849Macro? Micro? Wtf some economic class bs you speak of
>>110090771remember when we hyped up Z Turbo being able to make anime? that was some insane hype
>>110091026v5 full is probably a gimped quant version of the real NovelAI v5, according to Kurumuz v5 works at 4k resolution natively and what we are getting here is always running at 1k.
>>110091155I could've told you the same, the first month the outputs where much better than what they are now. But still, its the best model by far.
>>110091026google? seriously?
>>110089385Bg characters sharing the same face and those 2 penguins being literally identical copies just one scaled down is kinda sus for a good base model tbqh.
I'm dipping my hand at training loras for the first time. Is it bad to have a majorly monochrome dataset? I'm trying to train a particular doujin artist style
>>110091190If you want monohrome output then go for it.
>>110091213ok you retarded faggot you clearly understood my question yet you choose to act like a quippy redditor instead of answering me.Hark, faggot: can I use say, 80% monocrhome, 20% normal colored and still train a proper lora?
>>110091223*yawn* keep the insults coming anon, my wife rails on me daily, one more line aint gonna touch this tigerYou learn to make good loras by making loras, nobody gets it first try, nobody can give you the exact config, you'll probably end up redoing this one with different tagging or config anyway, that's normal.anyway, you want color versatility? tag monochrome and any dan booru sub variants if your dataset has them and tag the colored stuff by how it's drawn, anime coloring, flat color, high contrast, painterly, whatever applies, or leave untagged if it's the style you want the model to learn, imagine gaslighting the model.tag as specific as you possibly can, use a sane lora config, train 2.5k steps saving every 250 so you can compare.
>>110091320so you CAN be a helpful faggot when you want to! I'm asking here because last time I trained a lora was about when the NAI leak happened and people were molesting hypernetworks
>>110088619We are both oldfags here so cut the shit, MOSS is a mediocre model, same tier as Boogu or Ernie, maybe lower. You're simping because it recognizes a few gacha girls, that's not exactly a high bar. Even Flux Schnell, Kontext and ZiT can make decent furniture and chibis, come on man. Pic related is Flux Kontext btw, you are so starved for novelty you will praise anything new that drops.
>>110088379please use a better model than sdxl bro
>>110091450I agree it's mediocre, but sometimes bad is good - like "Plan 9 From Outer Space" bad. I'll know soon enough if it can be trained, or if it just falls apart. I get good anime out of krea2. Maybe at some point the Chinese will give us MiniMax-M3.Here's a blue-balls-board safe Shun aimation
>>110091450But anonnie they are making a dedicated anime model as we speak. Just wait two more weeks.
>>110091793>no e621 tags>no artist stylesDOA
>>110091793Who is posting this? is it from the openmoss team or some redditor finetuner using gpt image as cover images for his model.
>>110091824It's real, there is an ama on the stable diffusion subreddit.
Why even waste time on a shitty 5B model, its just gonna be a bit better than anima, but nowhere as good as NAI v5
Still fucking crazy how the entirety of imagegen still owes the NAI leak that happened years ago. Some retard literally burned a zeroday on github for it
>>110091824One of the moss team devs https://www.reddit.com/r/StableDiffusion/s/6WtpXf6qySThey are not satisfied with the current anime situation and they only used a tiny fraction of the dataset. Local will be saved, trust the plan
>>110091875He should've saved it for v5 dang, maybe some chinese chads will replicate or reverse engineer it.
>>110091882>>110091793Friendly reminder that ComfyOrg still didn't announce this great model even 24 HOURS after release, but sure had time for Ernie, Flux 3, Hunyuan and whatever other slopCOMFY GRIFTER YOU AND AND TURDY DESERVE THE WORST
Turdycord link?
I've done LoRAs (and seen other peoples LoRAs) on artist styles and none of them are as good as native nai v5. So if MOSS fucks this up it surely wont survive civitai's collective attention span.
>>110091921The only place you will find tdrusell is inside his lamborghini
>>110090592>needs a lora for KonataNo thanks.
>>110091944That collective attention span works both ways.
Anima stays king forever, you can train it on a consumer gpu.
>>110091921Yeah he rarely showed up there. He's trying to keep his hands clean, so to speak. Don't sniff the used rubber though.
>>110092002Huh? Dont get what you're going at, im just saying that the community is segmented enough as it is, no one is doing these big fine tunes like in the SDXL days because we dont even have a consesus on where to focus.
>>110091349>I'm asking here because last time I trained a lora was about when the NAI leak happened and people were molesting hypernetworksWho asked? Why acting like a tough guy with forced insults and then switch to little bitch mode unloading your life story that nobody asked for. Cringe.Go back to discord, you little troon.
>>110092073DESU I'd like to see someone stuff the SOTA model copy and draw routine into a local-sized model, and then basically you could ask "draw me the same thing but...", and no need to train on anime, just show it an anime picture you like.
>>110090698
Shindol, Nakamura Regura, Oltlo, Butcha-U, Soraoraora, Asanagi, Dagasi, etc. These are artists MOSS Anime® should generate just by prompt or it might aswell not release.
>>110092148>liking all those nigger artstyleskysMOSS anime needs only homunculus, takayaki, saitou masatsugu, kurono mitsuki, muririn, moekibara fumitake, etc.
>>110092320You sound like chris chan desu
>>110091882Finally, good to see an anime finetune actually using flux2VAE to its full potential. Every other model that released on this VAE got slopped into oblivion to the point where it makes me think Flux VAE itselfs is a big chunk of the training problem, especially considering how shit flux's own models are given their param count.
>>110092320Most of those look like generic anime tho?
>>110092418>generic animeThat's called cute and moe. The fans of hags with tiny heads, chest tumors, and landwhale hips can go back to western artstyles. Not anime.
>>110092513Nothing wrong with moe or any of the artists you mentioned, but you're autistic for starting a "my artist good your artist bad" shitfling. I find the artists I mentioned extremely distinctive and unique from each other, unlike those you mentioned that might aswell be generated randomly from any AI.
>>110091162>the first month the outputs where much better than what they are nowYou mean the bugged weights? You can recreate it, V5 uses a complexity tag>low complexity, medium complexity, high complexity, ultra complexityLow will give you very simple symbols and look more web comic-y, while ultra looks like that first month of V5's release
>>110091882They're from a university, wait till they have Ilya getting raped and her tight child pussy dripping with cum in their dataset
Holly fuck I love InvokeAI so much, the best thing that could happen to an img2img fag, Krita is autistic.
>15 hours 48 minutes>108/26/Who left the door open and let the schizos enter
>>110092588kek. imagine curating dataset captions and having your professor walk in on you
>>110089332ctat>>110090698falt>>110091190bnuy
>>110087668It must be both hot and frustrating to want to fuck a girl like this who has no holes, leaving you to only thigh fuck her.>>110092588Wasn't Noob —dan and e62x dataset—a uni project too?
>>110092667>nyoo dont discuss a new model
>>110092588moss can already do the cups lol
>>110086767https://pelune.studio/
>>110091190Monochrome data is fine so long as you tag it as such to make sure the robot knows it's not an inherent feature.
>>110086767Any anons interested in me making Hisoka gens for them? /a/ seemed unexpectedly uninterested or even outright hostile (muh AI slop) to it.
>>110091190>I'm trying to train a particular doujin artist styleGood luck with that, anon. Lemme know how it works out for you. I'd love to do the same for TRUMP/Masakado (Team Shuffle) but wouldn't even know where to start, especially since I don't gen locally. And yes, someone did a LoRA for him but it's evidently not trained on his yaoi works so not all that useful for my purposes.
>>110094706Okies :3And yet they consume any slop seasonal anime, ironic.
>>110093710cui_ veredict? moss or krea 2? For me krea 2
>>110094951>Okies :3That a "Yes, I want some gens", anon? If so, elaborate.>And yet they consume any slop seasonal anime, ironic.Well, it's not called "tranime" for nothing and trannies are notoriously anti-AI by and large so following that logic it shouldn't have come as too much of a surprise that I'd find backlash there, I reckon. It's more so the apathy that surprised me. Maybe I'm just that late to the party and everyone and their dog already moved one.
>>110095038>>110094706you sound like one fragile snowflake
>>110095205Why?
>>110091486that was wai illustrious trained on some twitter genners's genshttps://x.com/rin_illust_sub
>>110091882Another anime base model? so controlnet is some relic of the past? same deal with anime edit models i guess. i don't get it, don't see how doing Ibuki in Anima is going to look any different than doing her in MOSS.Illustrious, Anima, Neta Lumina, Newbie, Mugen, Mage Trial, please stop with this faggotry.
>>110095272Way better vae, qwen vae is horrid, and anima has to small of a text encoter.
>>110095205???You an actual tranny? Or what set you off about that comment? I don't get it.
>>110092633Any tips for img2img? I'm still using Invoke for prompting mostlyWhat sucks is that it doesn't let you upscale with Anima
>>110095328Grew up on dbz, sakura cc and samurai x through a tube tv and somehow survived without an ultra sharp vaemust be a zoomer thing then ¯_(ツ)_/¯
>>110092633>>110095372Do either of you do any smut img2img stuff? Got a technical question of sorts.
With the help of dan booru and gemini to set up the panels and scene i want to adapt light novels to colored mangas using ai diffusion, how can i do that?
>>110095755Not to shit on your idea anon but you are like 5 years too early to be asking this at least if you want it to feel immersive and enjoyable instead of a second job.
How would you guys go about making some money with this. I have some extreme fetishes, but I'm just so disappointed with local generation and Anima. Would really paying for novelAI make a return on investment eventually? My fetish involves loli, and pregnancy.Also, does NovelAI let you do like a lore from a certain artist or something like that, in case it's not in their training?
>>110095907These days it's easier taking commissions for loras, there's no shortage of idiots who will pay for a lora even if it's overbaked garbage. Use your NAI sub to scrape an easy dataset if you wanna go full grifter slopper, your clients will never know.
>>110094706i tracked you down.hisoka height mogging the demon manlet hiei if you're still down.
>>110095979You got it, sir.(Didn't we meet on /degen/ though? Or /slop/? Or am I totally off-base?)
>>110095907>How would you guys go about making some money with this.Produce good stuff. I get regular messages asking about commissions and donations.I turn them all down.
>>110096080yeah, shlop.
>>110096241Well, good job finding me then. Coulda just replied to my i2ibeggar request post though, unless you missed me pointing out last time that that was a way to contact me.Anyhow, sorry for the wait, super busy. Here's the first attempt at it (raw). Lemme know if you want more and if so, what you'd like to see changed/genned instead.
>>110096305thnks. just the one is funny enough thank ya
>>110096388Well, happy to have helped then, anon! Have a nice evenin'/good day/good night.
>>110095958>My fetish involves loli, and pregnancy.gen it with cuteyukimix-adorable (sd 1.5) and then i2i it in a newer model to upscale it?
>>110096305>>110094706cute
>>110097345Thanks. Feel free to request something of your own, these tokens are burning holes in my pockets.
i want to finish a few images but it's so much easier to just gen stuff and pick a bunch of images to inpaint/fix later
>>110097361Imagine her feeding you, too. Bliss ToT
>>110071944can't explain why but those frills have very illustrious smell to them
Did comfy added moss image to the templates?
Never tried image gen before, installed comfyui and anime but I keep on getting no faces on my slop, what gives?
why cant monstergirls be real bros
>>110098461>We're gunna need ze Heavy Flamer for zis one, Hans.
>>110098461yamamer is too cute to be a monster girlhttps://files.catbox.moe/68vouv.jpgkappa and tengu are the monstrous oneshttps://files.catbox.moe/9kf4ms.jpg
>>110098416Comfy is trash just use forge neo>>110099243Artist style?
>>110100626>Artist style?photo (medium), figure, photorealisticSeems to be scooping mega_man_(classic) for some styling?
moss lora test
fux2vae is poisoned just like their models.moss anime will be a new neta lumina
>>110092513>moefag complains when every single illustrious finetune already spits that generic slop by default
>>110100666How it handles your lora + full body + background?
2026 and illustrious noob still show up on a ctrl+f in these threads
moefags are hentai noobs
>>110100991
>>110101033legs too long
>>110098461Monstergirls aren't buying you a GPU tho
>>110101049>>110101033Appreciate it. Thoughts on how your lora performeded with Moss compared to Krea? I'm having a hedgehog's dilemma with Moss right now.
>>110098134you would be wrong because it's anima
>>110101099On krea2 it didn't blur godzilla
>>110101539Mmm that could go either way, blurring the godzilla might actually give the image a sense of depth
>>110100016is this krea2 with a lora?
Pre anima rakun was goat, he was building a good style.
This is rakun on anima, i don't like it
>>110101705Yes https://litter.catbox.moe/9jhndv0se2qd2jls.png
>>110101648Moss is going to be the better once the anime finetune is finish, Krea 2 is stuck on the turbo model and while it's fast and good it takes the shortest path when it comes to x, y and style (<--important) making it sloped.
coming back after a long break, should I check how far Anima went or are these newer models I haven't heard before (Krea, MOSS) worth looking into instead? Do the newer ones still need loras for niche characters?
>>110101915
>>110101950which of those is "it"?
>>110101915.
>>110101915>>110101962..
>>110101915Anima for danbooru artists or super niche charactersMOSS for anime images where you also want something a bit more complicated or want aspects of realism. MOSS is also completely uncensored and can be used for hentaiKrea2 for complex prompts. Probably better than MOSS in most cases but doesn't know nearly as many anime characters
>>110102063so, basically I'd have to try them all?
>>110102081pretty much. the MOSS devs said they're working on an anime model and that would most likely replace anima but on the other hand the z image team said the same thing and it never happened
>>110102091thx, guess I'll try anima for now and check moss sometime later.
>>110101915https://huggingface.co/RicemanT/MageTrail
>4chan's official anime diffusion general>48 images>6 reddit screenshots, 1 gemini screenshot
>>110102280that sounds still being wip
>>110102300who is this 4chan?
>>110102320
Is julien here? I have a question
>>110102791Tell me
>>110102410Moss?
>>110102988>bean and broccoli for breakfasti know shes bri*ish but thats a bit too much innit?
>>110102982>>110102988Moss?
>>110102993>breakfastwho said anything about breakfast?
>>110102993Ketotard here: i always eat broccoli for breakfast
>>110101750>>110101759Did the blog factory explode?
>>110102580Intentional prompt detail or unintended model artifact?
>>110103048it was unprompted and i couldnt tell if it was an actual error or not since its possible to hold a dagger between her fingers like that
>>110103066Thanks, one more ask, could you gen this same prompt on Krea 2 so we can tell if that melted ornament is a VAE thing or a low parameter issue? Or just share the prompt and will gen it myself to test.
>>110086767https://github.com/KohakuBlueleaf/KohakuMossTrainer
>>110103085the ornament wasnt prompted for. heres krea2 version with the same (anima) prompt but if i add a detail about the necklace in the moss prompt it might end up better
>>110103122This is the Krea 2 version???? ToT
>>110103102Maybe we should reorganize the OP in the following way:>UIsAnistudio.......>[model A][Hugging Face/CivitAI page of model A][Tag explorer of model A][Link to model A's trainer]>[model B][Hugging Face/CivitAI page of model B][Tag explorer of model B][Link to model B's trainer]Etc, etc.
>>110103152well kroma 3.1 version lol
>>110102998nai
>>110103355:^)
>>110101759Anima killed off a lot of good SDXL genners, they traded aesthetics for a better VAE. Not sure if that's on tdrusell's dataset curation or the model itself tho.
>>110103678cnnuy
>>110102982Gen Enma Ai, gen Enma Ai!!
is teto good for you?
>>110104231Imo genners who main teto and miku are the ones with more skill and experience.
V4.5 eris greyrat
>>110104345comparisson with other versions?
>>110104366Last thread there was a v5eris
>>110104375mkay, for easier comparisson>>110084325
>>110104398>>110104345
>>110104231teto cured my depression
>>110104375>>110104398>>110104426I think you want the strongest mushoku style, you should try V5 curated. Especially since you are genning for blue board.Anime screenshot styles got diluted in V5 full.
>>110104426V5 forgot the hairband and all fingers of the hand have the same length
>>110104426v5 looks actually close to the show's style. 4.5 looks like a generic screenshot style.
I want local nai...
>>110104472i want local dalle but we can't have everything can we
are the moss haters actually leveraging the json prompting?
>>110104624????we love moss here but using nodes 2.0 for bbox its a boomer
>>110104426which one got the canonical bust size correct?
>>110104452opinionated take, neither prompted for it and there's no reason Eris would be wearing a hairband to go to sleep and in the anime she definitely wasn't wearing one when she was about to sleep.
show me a good moss/krea gen
>>110105145Moss:>>110103249 >>110101033Krea:>>110103122 >>110101539
>>110105248he said "good"
>>110104624I don't know but i have been testig some anime loras for MOSS from different authors and it's hit or miss, it's like the vibe of the style, not the style itself It gives me SD 1.5 flashbacks
>>110105371>>110104659Why shouldn't I create a workflow that gives MOSS output a subtle, gentle pass through another model with better LoRA support? It seems like it would yield the best of both worlds.
>>110105272moss:> 110086993 >110088273 >110089385Krea 3: >110087136Krea 2: >110093710
>>110104182Aight. Even got the nose for you.
>>110105371not the best swarmUI test but still clear as day, krea (bottom pic) captures the style way better here
>>110105441This is an anime thread, crop the image and just leave the anime part
>>110105441the hell did it do to that poor cat though?
>>110105441>Homer's right hand has five fingers
>>110105441Lmao moss mogged krea for text.
at least it knows 2hou with no loras
>>110105396Dumb as hell. MOSS base model is straight ass, worse than basically everything else out there to be a decent first pass, Krea 2 wins by a mile. People are hyping MOSS because it knows a few anime characters, and even that's pretty surface level shaky. Example of a stupid hyped person that englobes what i'm saying: >>110105545
>>110105545krea2 can do some of the popular ones. not nearly as many as moss tho
Can it does iM@s though?
>>110105545That's no excuse for it to be a good base model, or even a good architecture to train on. They could've done the same thing with Z Image. And you'd probably have reached the same dumb conclusion.
why is lilbro in the anime thread complainin about models knowing anime
>>110105559So you are telling me that using the 3D stage studio of something like Artcraft and my usual models would yield better results and still alow for the same granular control over composition? What about text and text placement? I often use nano banana pro for those details, but I feel dirty because I want to do everything locally (ideally).
Kayoko wife
>>110101049legs not necessarily too long compared to the training data perhaps. I'm out of my depth but i know bikini often pose on tippy-toes. It makes you look taller and stand up nice and straight. Advice for photographers is also to make girls do this.
anima eats all my ram at an alarming rate.
>>110105596Yes, even you could stay with sdxl controlnet, microsoft florence and adetailer pipe for individual consecutive masks
>>110105396at that point you're just doing image2image
>>110105692yeah, but with way more control over composition, text and whatnot>>110105667that is what I thought, but I worry about text and text placementliberate me from the banana
>>110105707the second pass isn't going to turn character A into character B without distorting so much from your image you're not going to like it. Would be my elaborated guess on this matter.
>>110105579I wouldn't say it's dumb. Part of running local is you end up jumping to conclusions and daydreaming about what could work. Still downloading custom nodes trying to fix the grid problem on Qwen 2.1 VAE. Guess that makes me a dreamer more than a clown, like John Lennon said.
is it possible to generate images like (https://boards.4chan.org/vg/thread/584889659/zzz-zenless-zone-zero-general)(on the right) this using AI? My setup is using picrel and almost everything comes out as clear slop.How can I improve the output quality?
>>110105757That image was made with novelai v5full
>>110105423You know me too well...
>>110105588Loras trained on MOSS aren't looking that good in practice so far.
>>110105893Holy sexo who is this sex demon
>>110105918Kazusa (blue archive) in her official Halloween costume.
>>110105785I wish to gain this power...(without paying)thanks anon
>>110105931Thanks, is it from September 2025?(anima base dataset cutoff)
>>110105949Anima knows this costume, yes.
>>110105743Feel you anon, local model gave me attachment and trust issues.
>>110100230how do you make this quality??please teach me your method
Why are we back on booru era tags? What is going on?
>>110106646Turns out, booru tags are an efficient way to do some prompts.
>>110104345I've got a krea2 LoRA for Mushoku Tensei: https://huggingface.co/quarterturn/krea2-mushoku-tenseiNext LoRA will be Gabriel Dropout. I found some shortcuts for removing duplicates from intro/outro stills, it's made the process a lot easier. Having qwen3.8-flash-next running fast and local has helped a lot, it's really smart about this stuff.
>>110105596>>110105667What are you guys talking about, I only know how to prompt Anima
>>110105559Moss just gives me a nostalgic SD 1.5 feel, the good and the bad parts. But maybe it can be finetuned into a new illustrious.
>>110102986Krea 2 w/lora
I've been out for a while, is Krea good now? Like, for making Anime nsfw (loli, booru artist styles) basically a better Anima model, or should I stick with Anima. I'm using Invokeapp btw
>>110107089No, use moss sft for that, it's uncensored out of the box and you don't need cope loras just to generate pussy. For art styles isn't that great and character knowledge is so-so, not horrid but not great either.
>>110107188What about coherence/prompt adherence or whatever is supposed to be called that is the important and relevant part, because characters and styles can be trained right? Or first impression is that it sucks at that?
>>110107215It can do 4 characters without prompt bleeding just fine if that means anything to you.
>>110107089Yes, Krea is good but>loli>booru artist styles>charactersLOL no.
>>110106734Very well, indulge me and render Eris Greyrat. I do so enjoy watching little experiments reveal their differences. If you want the prompt NAI hides it away via stealth metadata, If you doubt my word, search for NovelAI Inspect on Google and verify it yourself, not that I'd lie to you.>>110107014Amusing, even megacorporations deserve a touch of cuteness.>>110106754I do hope the Chinese take it upon themselves to produce a proper finetune, otherwise *drinks her tea*... training such a thing under today's prices would be a rather tedious act of self destruction, wouldn't you agree?
>>110107260Nice seeing gens using moss image instead of all the doomposting.
>>110107260
>>110107089>I've been out for a whileYou are better off staying out a bit longer, we are stuck in a plateau right now.
>>110107370Prove it.
finally setup comfyui with anima to generate my own fap materials, this is going to be me in a couple monthsridiculous how the boorus barely have images that make me hard as diamond and cum buckets
>>110107280only hopeposting, trust the plan
How is moss better than anima?
>>110107417i barely look at anything that isn't my gens anymore. sometimes a good artist will release a book i like but that's about it.
>>110107418they're not gonna train it on degenerate stuff like actual sex content, futa etc, i have 0 expectations for anything that doesn't advertise itself as a porn model first
>>110107231No. You shall have Norn.
>>110107451The model is already uncensored, why do you think they won't do it.
>>110107418>>110107418They are using AI generated dataset for both anime and realism stuff. I did a couple of 2,3,4 anime girls myself, and the GPT Image slop was obvious as hell. Also, they never bring up Danbooru once in their MOSScord so that's probably not even their source. It smells like some random generalist AI dataset, not a proper booru scrape.
>>110107605gg/s838SFu8fTheir Discord is public, go read it yourself. The devs are clueless about what Danbooru is, what anime is, and what counts as slop.
>>110107490Lmao stiff as hell
>>110107784Then we have to teach them.
>>110107784They are from a research institute, who in their right mind would publicly declare they are scraping a dataset?
>>110107902Where is that hand going...
>>110107784this is not the discord link
>>110108077>>110108077>>110108077>>110093086added>>110103102fuck, I forgot this. remind me in next thread
>>110107796That's what she said!
