Responsible And Diligent Sensei EditionPrevious Thread: >>>>8815821>LOCAL UIreForge: 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
>Why are you doing [thing]Why not, I am just that bored so anything goes I guess>Why is so [thing] Because it's me>Are you going to do [thing]Maybe I will maybe I won't
>>8827656can you make [thing] more [thing]?
>>8827658I forgor about that one, fuck
This one didn't came out quite like I want it, boobs look weird
Have you seen any of the epstein files? crazy stuff , seeing and reading some of those made me remember all of those really old and obscure vocaloids songs, not cool
I was blessed with this seed the other day, it looked funny so even if I had a little too much gens like this I decided to complete it anyway
By the way, here is the gen of the other day, you know after doing more of those I think they are not that of a big deal anymoreIt's a little fried since I didn't intend to share it but eh, whatever, world's ending>https://files.catbox.moe/mm1l50.png
this is anima thread now
I have done so much shit over the past two years on this hobby chat, kind of nutsHope the new models once complete trained and full tuned will be great, as you may know I am more than fine with SDXL but new toys are always welcome
sex with JKs
Despite everything, my favorite shitmix is still 102d custom, r3mix being second place and bakaschizo "finetune" third place for the rad styles it can do (when it works after layers and layers of hacks)
I love 1girl so much, you have no idea, I thought about going back to posting on /e/ but it's full of very autistic childs nowadays, not worth itPeak /e/ was back then on the golden days of NAI v3 when the other baker was around, I wonder where he is now
what's the best online UI right now, paid or not? (apart from NovelAI)
>>8827677He, allegedly, is still sometimes in /hdg/ Last time I saw it was when he said that he closed the /e/ tab and left as soon as he made his last post.
>>8827677the meltdown over the past two weeks or whatever on /e/ has been entertaining. carebear schizos have been outed as being control freaks who are only okay with things when they're the ones doing it and a few of the remaining "regular posters" have gone mask off schizo-mode when anyone says something they don't like.meanwhile /adt/ is completely fucking dead with the last two threads being schizo highjacks and the only real poster left being a specific nai user, with occasional schizos that "migrated" to edg making light appearances.
What's better than 1girl? 2 of themI love yurishit and futashit and really really want a place to post about it but for some reason, those two always attract a very particular set of autists that ruin the fun for everyone else, I don't want to see girls with giga cocks or balls, I want to see girls make out and/or imply that they are fucking the other one
>>8827684I also have thought about going back to /adt/ but I hate forcing myself to do sfw when I am clearly not in the mood of doing it and "barely" sfw would get me banned on this dogshit website
Speaking of it, I still want a "low level" bestiality thread, where the main topic (and focus) are the girls taking animal cocks>https://files.catbox.moe/j6se2o.png>https://files.catbox.moe/qxsxzf.png>https://files.catbox.moe/kpqk03.png
>>8827691Based. I personally feel us swinebros get no respect in this field. It's churls, orcs, gobbos, minotaurs, etc. But not enough ugly piggus.https://files.catbox.moe/w658jn.png
>>8827684>the meltdown over the past two weeks or whatever on /e/ has been entertaining. carebear schizos have been outed as being control freaks who are only okay with things when they're the ones doing it and a few of the remaining "regular posters" have gone mask off schizo-mode when anyone says something they don't like.I feel like it's mostly the IA schizo and maybe 1-2 of his most loyal minions. Most of the other slop posters don't really give a shit what happens around.
>>8827665>seeing and reading some of those made me remember all of those really old and obscure vocaloids songs, not coolWhat's the connection there?
>>8827684/adt/ schizo highjacks?
>>8827691How about >>>/trash/wfag? It's not a pure AI thread, but gens were welcome when I last looked at it.
it's so over
Loving Anima so far but I'm having trouble with style consistency. It's like I either gotta put in an artist as a tag, or just get random shit every time?I don't want it to look hand drawn, I just want it to look crisp and detailed in every image, like default Illustrious type style. anyone know how I can prompt that?
>>8827783>default Illustrious type styleYou better ask /edg/ then
>>8827783>like default Illustrious type styleDefault illustrious is like this too, you're thinking of WAI or some other mix. It's similar to having an artist prompted at all times with low emphasis, someone like dikko. You can also run a low-strength style lora.
>>8827783models should be unusable without an artist tag. fuck you slop faggots
>>8827783>typing two artist tags is so fucking hard, I want some indian slop modelmaker to handle it for me. i want it shiny and butiful detiled!!Wrong thread.
Even with NL I can't get the guy to grab the breasts in a way that presses them against his dick.
>>8827790>>8827788>nooo you're not allowed to goon until you can name 853 different hentai artists like i can, if u dont follow and track of them ur not allowed to goon >:( you faggots are fucking weird>>8827786Great, I'll try it out
>>8827793you're a fucking monkey
>>8827792not perfect, but still, it's a prompt issue. think of another way to approach what you want to see and write it out.
>>8827796eh, not terrible for a first past, i'll try messing with the tags a bit more
>>8827652any chance for op box
>>8827800it's NAI
Reminder that /e/, /g/, /hdg/, whatever refugees are not welcome here. If you weren't part of #001, you have to go back.
>>8827783the model not having a baked in default style is a good thing, slop-kunliterally just put together an artist mix for a style you like or downloadhttps://civitai.com/models/723360/ai-styles-dump-animaillustriousrouweinoob?modelVersionId=2663471
>>8827797I want to nibble on those nipples
>>8827816it absolutely must have a shiny brown innate style that ruins the model, we saars love that shit
I'm not downloading comfy again, no matter how much you talk about anima
>>8827823enjoy getting left behind, gramps
gradio will adapt
do you guys remember when naiggers first posted about text being possible. this kinda feels like that
>>8827830https://files.catbox.moe/i8ibgd.png
>>8827833>girl is looking straight at the viewer, she is indoors and the background is blurry. There is sunlight streaming through the window, and her hand is on her chin with a finger on her lips. Her large breasts are visible, and she has puffy nipples. Across her breasts in large, permanent marker is the wordsI hate natural language
>>8827836>mfw I can just type a sentence and it worksYeah god it really sucks
>>8827836>I hate natural languagewhy
>>8827736Well, some of those songs are related to allegaly some real cases of japan but are on the same level as some things described on the files
>>8827778Uuuuggggggghhhhhhhhhh fine, I'll gather some images later
>>8827836just mix tags with NL?just use tags?
>>8827744That thread is surprisingly active wtf
>>8827830>>8827833> three years later
>>8827802Based but the general spiritually ended at #10 and since then it's just been the network throwing tourists and dregs at it hoping for a comeback.https://files.catbox.moe/kpdskl.png
My biggest problem with Anima is that it can't really do "textures", as in artist styles that have more detailed skin textures or shading that is neither super soft or completely flat, which I'm guessing is the consequence of training on 512x and so I presume that will be fixed for the full release if they train on 1024x
>>8827891Do you have some artist examples?
How to lora anima? google search gives these reposhttps://github.com/bluvoll/diffusion-pipe/tree/mainhttps://github.com/FHfanshu/Anima_Trainerone is bluvoll one is chinese. not exactly confidence inspiring
>>8827891>model looks like slopis this the time for sdxl refiner to make a comeback??
>>8827898>sdxl refinerNAIsissies are going to have a field day now..
>>8827893Yeah, stuff like Yunsang or nanaken nana and just today I tried MireiThey end up looking more flat than they should because I'm guessing the image is too compressed as to fit 512x so they detail is lost when training
>>8827901>theythe
>>8827894https://github.com/kohya-ss/sd-scripts/pull/2260It's getting added to kohya-ss
>>8827901Could also be because there's a huge style difference between their older and newer work.
>>8827901>>8827909yeah every next image is in different style
anyone here tried training an anima lora on diffusion pipe with 10gb or less vram? i keep ooming
>>8827801Box it anyways.
>>8827905duongve is the chromayume/netayume/illumiyume grifter, he just vibecoded bluvoll's implementation into sd-scripts >>8827911use year tags
>>8827846Which songs are you thinking of?
>>8827916so bluvoll vibecoded an implementation that was further vibecoded into sd-scripts? where is the machina guy when you need him?
>>8827916use year tags and then pretend they work with artists rather than just being a generic style bias towards modern slop skinshine
let's face it, we need loras
let's face it, we need wai shitmix
>>8827936https://civitai.com/models/2361209/perfectrsbmix-anima-aio-sfw-nsfw?modelVersionId=2664883you called?
>>8827938absolutely love it
>>8827938waow!
>>8827849pay debnts
>>8827938horrific
>>8827938>Now we don't need use improvements tags>Remember use masterpiece, best quality, score_9 unless>Still is working as preview, tags or natural lenguage
>>8827938https://civitai.com/images/120259466home
>>8827950god. and fucking brown skinned degenerates will praise this shit.
>>8827913answer my question... NOW!
>>8827909>>8827911>>8827916I've already tried prompting for year and using "newest", it has barely impacted anything
>>8827938>>8827950>AOM sloppa Anima edition
>>8827961you should buy a new gpu NOW
>>8827966this. nvidia deserves our support
>>8827950>>8827938why tf do i need to log in to see sensitive content when porn sites don't have that shit
>>8827974Because civitai is not (allegedly) a porn site
>>8827921There is a few of them, this one was the first one that came into my mind>https://youtu.be/I1mOeAtPkgk?si=OhMMikm0Urx0sbUV
>>8827980unc...
anima feels a lot like nai in how it can't do a single style properly and forces you to mix. (not that mixing works all that well either)
>>8827982Yes yes I know, to be honest I don't want to comment anything about it but he asked for it so lolBack to 1girl and sugoi hentai
Let's all just have happy thoughts
>>8827980I see, thank you.
>>8827964yes well >>8827932
Hope texts gets better as well. would rather not switch to qwen for that
https://www.reddit.com/r/StableDiffusion/comments/1qyk4fd/anima_2b_style_explorer_visual_database_of_900/https://thetacursed.github.io/Anima-Style-Explorer/
>>8828007Seems pointless given how low image-count artists work even better than noob. There's gotta be tens of thousands.
>>8828013yeah he kind of just stopped at 700 for some reasonI use a lot of artists that are even below 200
>>8828013Not in his examples, it's all slopped up with quality tags.
https://files.catbox.moe/talhli.png
since kohyAss is used on the gui version, will updating kohya alone fuck it up?
>>8828057"sometimes"yes, it's as vague as that. also if you're trying to update kohya directly to access new training implementations you'll be locked to having to use kohya directly and not using the GUI at all(since the GUI won't just magically update to include all of the specific new shit).
where can i find illustrious loras for the character uni by arekishi? civitai is gay and can no longer find the rentry dump where some anon linked one
>>8828065illustrious lora rentry is in the /hdg/ OP
>>8828078ty, but what's the difference between hgg and hdg? isn't it about the same stuff?
>>8828079/hgg/ is cool, /hdg/ isn't
>>8828013idka lot of artists are kinda ass
Anima full release when?
is nobody really using chenkin reflow? seems okayish.
>>8828158I am going to test it probably tomorrow since reforge has support for it but I am keeping my expectations really low
>>8828158It's definitely the best SDXL model we have but anima is a new arch and despite it being undertrained and mostly trained at 512x512 so far it's a bit more interesting to play around with and they dropped around the same time so it got overshadowed
>>8828158...why? anima exists
>>8828158i downloaded it the first day it was uploaded to huggingface and completely forgot about it because of anima tbqh, i'll give a try.
>>8828158Quick 0 effort raw upscaled gen, not bad I guess
why is anima so shit at artist mixing, it combines the worst aspects of the artists
No refunds, chudai.
>>8828171LLM TE
>>8828165Alright, slightly more effort this time, also tested how well inpainting work and it does fine, maybe I'll use it a bit more to really see the extent of the updated dataset used but I really really hate using the suggested samplers
>>8828176Style wise doesn't seem to be much different from the goat 102d custom, it's more blurry and sketchy but I bet all my horses that it's because of the suggested gen config (quality tags, sampler, CFG, scheduler and steps)
>>8828177>102d customBet you didn't even mix it with sekret 29+1, coward.
>>8828182Of course not, I am not a homo
>>8828185Your gens suggest otherwise, unc.
it's insane how boring 102d custom is. might as well be wai, but without brown gloss.
>>8828188Brown hands shitmix. Please understand.
Is it just impossible to control what 2 different characters are doing with Anima or am I retarded?
>>8828188Holy TRVKE
>>8828199latter
>>8828199you just need to prompt the way you normally would with danbooru tags. when you think you have everything ready, translate them into natural language, but don’t delete the booru tags. So I’d say booru tags + NLP is the best way to use Anima. Remember to specify who is where and who is doing what. futa warning: https://litter.catbox.moe/1biift1c7w24e3ss.png
>>8828209>No metadata found.You're not obligated to share, but if you're not willing to help, why even bother typing all this shit? It's useless.
Based concern troll.
any recommendations for artists that draw hot black women? (not a shitpost)
>>8828221That nigga that was in every aco-like mix back in the day. Uh.. draws a lot of fantasy shit. Incase. There we go. Might also try mossacanibus or whatever they're called.
>>8828226mossacannibalis looks good, thanks.
>>8828221Is there something to black (anime) women other than changing the skin color?
Personally, I think it's time we address the gacha slut menace and it's long term negative health ramification on ojisans.https://files.catbox.moe/n26tcn.png
snail shell - grisa alt color
>>8827894https://github.com/tdrussell/diffusion-pipediffusion pipe dev is also dev for anima so its gonna be the most accurate implementation
>>8828250I want to rest my head on those soft looking breasts
>>8828244glossy, lickable eyelids and lips with heavy makeup
>>8828256>lips with heavy makeup>>>>>>>>>>>>>>/aco/
I dont know if this is the right thread for it but im asking anyway. the artist of this pic used to have a large rule34 collection online but i cant find it anymore. artist is mino_dev. it cant possible have all been nuked. anyone know where i can find it.
>>8828268have you tried using wayback machine for the artist tag on r34?https://web.archive.org/
>>8828274nevermind just tested it myself and r34 is excluded kek
>>8828275>we need to preserve the internet>NO NOT THAT INTERNETDeath to puritan subhumans.
this, we need to preserve all the r34 comments for the future of the vvhite race.
>Nooo you have to preserve my favorite porn for me forever for free, I can't be bothered to download things I like.why are millennials like this?
>>8828284but the hecking PSArinos about what you post on the epic interwebz being there fivever???
>>8828281yea pretty silly
>>8828268Did he kill his own porn to not affect the game?
>>882831516ch is no excuse to not upscale your gens, lil' bro
>>8828319nah i'm good
>>8828319anima shit itself pretty badly when you try to upscale
>>8828275>>8828281r34 probably excluded itself with robots.txt or something
>>8828322Yeah, but check last thread. It's not quite when you try to upscale, it's when the model works at too high resolution. You can still do a 1.3x or so, or you can use tiled upscaling where each tile is only 1024 and go as high as you want.Although for the issues anon probably meant in >>8828319, all it would take is inpainting a few key areas.
>>8828327Ah it's not actually last thread it's >>8826986kys splitfag stop rebaking this shit
>>8828327What are the inpaint settings, because when I try it the results are pretty blurry
>>8828319Yes it is, Hires. fix was nothing more than a cope for the compressed VAE and a way to maybe unfuck the details. There's no need to upscale when the base image is good in the first place.
>>8828331forgot my slop
>>8828333>image is good>16 ch vae>vaedo you have eyes? unless it's pixel space you'll have to upscale.
has anyone tried prompting for funky camera angles on AnimaI'd like to try some unconventional stuff but I don't even know how to start describing something like that
>>8828333Right, i'm sure a 512x512 pic will not look blurry at all if I zoom it to fit the screen on my 1440p monitor
>>8828348thankfully i found that typing "interesting angle" at the bottom helps with that somewhat. it's what i really missed in sdxl
>>8828349Don't zoom, tile multiple. Easy.
>>8828333This will be true once we get rid of VAE compression altogether, only halving it is sometimes not enough. Depends on the image, the usual issues of faces too far from the camera, fine details like zippers or jewelry, lace, etc.
damn
>>8828372holy kek
>>8828329>thread so good dumb mf has to blame his own schizophrenia to justify his stupiditylove to see it and welcome, I will be around more often so feel free to come and say hi
>>8828379I was just about to give you the inpaint settings
>inpaintlol, snake oiljust reroll seed
>>8828381I just wake up broski, no idea what are you yapping about, I just saw "splitfag stop rebaking this shit" and had to reply
>>8828379Based noodlefriend
>>8828384Just use an inpaint model. Mask and reroll seed of that section only.
>>8828000peebox?
>>8828217I just thought it was better than just calling him a retard tbqh. He can figure out exactly what to do if he keeps trying
>>8828398it's inpainted, chubby.>https://files.catbox.moe/x5ie3y.png
>>8828372miyubox?
>>8828404https://litter.catbox.moe/6ii3kr4yddjgohev.png
>>8828405based and thanks
>>8828401very nice, thanks
>>8828252>install le heckin' wsl2Fuck github, fuck developers, fuck operating systems and fuck python. why can it never be be download loratrainer.rar, unpack and run .exe?
>>8828401>(@ibukichi:0.5)Do tag weights even work on anima?
>>8828417they do
>qwen3 0.6b encoder Was Anima targeted specifically for saars with the shittiest gpus? 4b model would've been so much better, not even talking about higher llms
>>8828422it's this or sdxl. ain't no 20b aniqwen coming your way
>>8828422>4b model would've been so much bettergo ahead and explain why
>>8828424more b more bettergot it?
>>8828417They do, but artist mixing is kinda fucked.
>half feel like actually doing something>except my chair has been slowly breaking over the last two weeks and really fucking broke this past wednesday >had ordered a chair that was supposed to deliver same day, never came>didn't come the day after, either>cancelled because it was clearly lost in shipping, ordered another chair that was supposed to come today>nope it's coming tuesday>meanwhile I'm sitting here in a chair that's tilted 45 degrees backwards so I'm just sitting in a fucking V with no back support and doing anything productive is no fun at allI hate amazon so fucking much it's unreal. how does a logistics company fuck up logistics that badly.be sore to lik and scribscrub for moar blog
:skull:
we're so back, gradiobroshttps://github.com/Haoming02/sd-webui-forge-classic/pull/726
>>8828443I read the code, there's a virus in it.
>>8828443I will wait for my negro pancho to port it to reforge
>>8828405NTA, but thanks. To reproduce it consistently [spoiler]on NAI[/spoiler], I had to add "cum in container, cum bath" to the prompt.
>>8828431yeah I've tired mixing some myself and it just doesn't seem to workI think it gets confused because of the necessary @, so it only expects one artist tag and not multiple @s
>>8828472Why are you using NAI when Anima is the new SOTA?
>>8828476Too lazy to install Cumfy.
>try out my favorite artist shitmix on anima>it's a garbled messNot like I expected anything else, but I'm still disappointed
>>8828481there just has to be something wrong with how it parses multiple artist tagslike it's similar to when you use an artist tag without the @, it kind of just doesn't understand what to do and so it gives you something random
localfrens are now discovering the joy of artist mixing without CLIP
>>8828496qwen 8b would've understood that there are multiple artists listed together in a mix
>>8828500Is this achievable on anima?
>>8828496We have prompt scheduling though, which helps
>>8828481did you remember to @
>>8828499Consider the following: Very few images on boorus have multiple artists, and those that do are usually artist collaborations where each artist draws in their own style. That means that a sufficiently smart model with an intelligent text encoder will understand this and will not mix styles when it sees multiple artist tags. At best it might start drawing parts of your image in one style, and another part in another style. This is what the data distribution teaches it.Artist mixed images are simply out of distribution, unless you train on SDXL slop.
>>8828508Has anyone tried "face drawn by abc, body drawn by xyz" etc
>>8828511I highly doubt that will work. There are no such labels in the data, and even artist collaborations are likely just split by character or maybe background and characters, not parts of a single character. If you are very lucky, the model will generalize to "background drawn by ..." or maybe "asuka drawn by ...", but that's a very long shot and I'd be surprised.
>>8828511Doesn't work, it just mixes them image-wide like usual.You can still do it with regional prompter+controlnet, once we get one. I've tried that on noob, made a raita skeleton with a zankuro moeblob face. But it' not what we're looking for here.>>8828492>when you use an artist tag without the @In this case it's because without the @ you're prompting a different tag, the TE doesn't leak like old CLIP. Or perhaps not as much.
>>8828508>>8828511>>8828523What if we use a ConditioningAverage
what if we use gradio?
>what if we use a UI with even less options
>>8828524How did you make it this far and not pressed Generate before posting
>>8828527more like>what if we use UI with less snakeoil and realize that arch problems can't be fixed by chaining together tranny nodes
>>8828524My expectation is that this will often result in unpredictable results. It might sometimes work to some degree, at least if both artist tags have the same number of tokens. Doing it for a full prompt with multiple artists will be difficult.
>>8828536yeah I just tried it, it kind of just creates a mess that doesn't really resemble either artist tag
>>8828539like it's just kinda shitty or the result is a complete mess?
>>8828542it's just kinda shitty, like it doesn't break the gen, the style just feels like something almost entirely newI'll see if I can put together a comparison but my gens are slow so it might take a bit
Anima is weird. It doesn't seem to recognize tons of popular artists, and at the same time I think it's the first model that knows Hanabi (Starmine18) out of the box (which throws the anatomy out of the window though) https://litter.catbox.moe/g02msogpkkatierf.png
anima is really good but the lack of artist mixing is a bummer.
>>8828556>tons of popular artistsSuch as?I've only tried a few dozen so far but everything worked better than on noob even. Well, uglier but more accurate.
I don't think ConditioningAverage does anything meaningful
>>8828556It has recognized literally every artist I've tried that Noob also knows. Make sure you have @ in front of the tag, and convert underscores to spaces. Make sure you escape parentheses in Comfy. For artists that have different spellings between Danbooru and Gelbooru, use the Gelbooru spelling. All these things can completely fuck up the artist recognition if you do it wrong.
>>8828561this is it mixed normally>@hood \(james x\), @ahonokowhich looks pretty okay so maybe artist mixing isn't a complete disaster
>>8828562>It has recognized literally every artist I've tried that Noob also knows.recognize is a strong word for the fidelity that anima usually produces...
>>8828570What are you on about, it's pretty good in terms of replicating a single artist, it shits the bed when it comes to mixing.
>>8828557isn't anima just a good base model we can expect it to improve a lot later, like base pony vs autismmix
>>8828580Autismmix was a downgrade.We don't need mixes after SD 1.5.
>>8828579Don't know what artists you are prompting for, but it looks nothing like the ones I'm trying
>>8828580Probably not better, just easier to use. Thing is, most of the Pony ones were merges with other SDXL models. This is an entirely new arch so there's nothing to merge it with.
>>8828583which ones are you doingpost an example of your gen in anima and noob
does anima use booru tags, I hate nat language for anime prompting. Not going to even bother downloading it to try at these early stages yet if so
>>8828580Questionable comparison but yes you should expect anima to improve, it's still undertrained and most of its training was done at 512x512
>>8828595I am anon, and I am really bad at reading. I am so bad at reading in fact that I didn't even check the models huggingface page, and instead wrote this here post on 4chan.org.
>>8828595You can use purely tags if you want. NL helps for things that cannot be well described with tags, or for positioning people/stuff in the image.
>>8828595You can but the NL is incredibly powerful for positioning and you'll be grateful for it when you realise you don't have to roll a bunch of times to get what you want.
I spent 40 usd on a NSFW app just to make shitty hentai images of some girls I know.Fuck
>>8828612Go to /edg/, they'll love your stuff!
>>8828612That's sad anon, I would have made them for you for $35
>>8828612>I spent 40 usd on a NSFW appThe NAI shill had one job and he failed you, I'm sorry
What creative things are you guys doing with anima I realized that all I want to gen is /e/ and 1boy 1girl so the NLP shit doesn't actually help me do anything creative at all.
>>8828619Kinda the same, but the NL gives more control when it comes to poses and camera angles, and the model seems to be less retarded than SDXL finetunes about random everyday items.
>>8828612mega F
>>8828612fucking YIKES
anyone know any good settings or does the default recommendation from the huggingface pic work fine?
>>8828619why force yourself to gen shit you don't really want to? prompt what you want to and be happy if your model allows you to do it without spending much time
>>8828625It works fine, but adding "deviantart, ye-pop" to the negatives helps reduce the /aco/ bias.
Ok but can anima do anal_tugging
It can't do thinly veiled requests
Does anima do some weird colors/lighting ever other gen for you guys as well? Almost feels like early vpred
>>8828630I find just prompting an artist is enough, unless I go really heavy on quality tags.
do you need to use a period after a NL prompt or can you just add them to a comma list of tags?
>>8828658doesn't matterit doesn't even need to be correct english, you can write five paragraphs as one sentence with no punctuation. I think it's just picking out phrases
no model will ever mix as good as SDXL
>>8828612Anon...
>>8828443>>8828453p-panchito.. y-yonde kudasai.. *cough* k-kite.. kudasai
Any hope for training Anima loras as an 8GB VRAMlet?
since kohya gui is never ever on updating it's own.... gui, anyone tried the thing called ai toolkit for lora training?
What's the best online service I can use to just quickly add cumshots to pics I already have
lol
>>8828680my nigga playing tibiagood for him
>>8828000Whats the source on this?Not getting anything on image searches.
are furries going to add e621 to anima? how am i going to make females fuck animals without their dataset?
>>8828209anon you should have been generating with a combination of tags and a prompt for fucking years now, this should not be a novel concept
>>8828729what's was point of NL on CLIP models. anybody thinking their SDXL model understood NL is coping.
>>8828731it's still going to parse the words and associate them with what should be properly tagged and captioned data. at minimum it should reinforce your tags' intent
anima kinda sucks at bleeding characters if you're not using NLP
duh
>>8828743I meant like in a very basic 2 female characters + 1faceless man. Noob can do it better. Or maybe it was just because i was using characters with less than 90 images
Still trying to find a good way of using names in NLP
>>8828758Up your unc game
>>8828758You can do your usual 2girls , madoka, homura; and then "say girl on the left has pink hair and a smug smile".
>>8828782>smug smiletf is her problem?
She was meant to be the crescendo of the parade of the newly captured, but it seems the stagehands had had their fill of her before she was even revealed.
>>8828758That's obvious: names are very useful for describing who goes where and does what to who. What I'm more curious about instead is how are you supposed to prompt nameless non-characters you made up. They don't have names so how do you address them in the NLP portion of the prompt? You can get away with "the boy" and "the girl" if it's just 1boy, 1girl, but what if there are two girls?
>>8828792Even you can't turn it around, Sir Solangeanon. We're beyond redemption.
upboat my post about artist mixing plox broshttps://huggingface.co/circlestone-labs/Anima/discussions/44
>>8828804it cant be fixedeven google's nano banana or other equally large models cant mix stylesyou can at least get some mixing with loras but just raw mixing will always kina suck compared to XL
>>8828806damn
>>8828807you can also hope for a vibe transfer/ip adapter like feature which will help mix styles
>>8828808>vibe transfernaisissies.. be gentle.
>>8828808>vibe transfer/ip adapterDogshit slop snake oil, what we need are more generalist reference images that could be used for more than just styles, mainly for characters, but also for concepts (poses, items, etc).
>>8828804Fixing it would mean more or less retraining the model with CLIP instead
>>8828814Snakeoil means it doesn't do anything. IPAdapter works, it's just inaccurate and can't separate style, character and composition.
>>8828817>doesn't do anythingThere are two ways to achieve this:>actually do nothing>try to do something and fail so spectacularly nobody will ever use itIPA is the second. Both are snake oils, just that the second one is by choice rather than by nature.
>>8828820Have you tried the noob trained one? It's still pretty cool, I use it on occasion. For transferring weird background designs or as added style bias on top. I recall cathag mentioning it too a couple times.It's just not accurate enough to fully capture a style, and nowhere near enough for characters.
>>8828822>cathagDon't ever bring her up.
>her
we will never be graced by epstein levels of devotion to a single general endlessly highlighting our posts again, or the accompanying instructions to escort discord users off the board with blocks of conveniently deleted posts like someone was trying to tell us something>waow>why seethe>declineno one will ever wear that crown the samebut there must always bea highlights
*smooch*
>>8828808what we really need is this model to have capabilities of edit model>>8828806how did almighty and greatest folks at NovelAI corporation figured it out then? (damn, now that i think about it, they kinda didn't, give how bad novelai 4.5 looks)
>>8828840>how did almighty and greatest folks at NovelAI corporation figured it out then?Synthetic data from V3 with mixed artist styles.
>>8828793"the blonde"? whatever is different between them
>>8828840>NovelAIRecommends negpipping artist collaboration to make mixes work.
>>8828850can we get things like negpip and NAG working for Anima on comfyui? severe lack of snake oils
>>8828840>how did almighty and greatest folks at NovelAI corporation figured it out then?like you said, they didn't. Also I haven't used anima at all because I'm not bothering with cumrag shit but a few things to point out.First, you should be checking one artist at a time, no mixing. then you should add an artist and see what/how much changed. then swap to a different artist, then add an artist, etc. Point is, going straight to a mix doesn't really help or tell you anything.Second, the "mix" isn't exactly something tangible to begin with. Any and all artists are going to be heavily impaired when trying to just replicate from the model and not from a lora. And "mixing" them is nebulous and unpredictable until observed(read: you will have no idea what it looks like until you know what it looks like) and that's on a per model basis to begin with.either way the "solution" is that with anima at least you can just train a lora. of course, you're going to need to have enough "high quality" output images to form a dataset but surely if you care that much you've spent the time cleaning, fixing errors and unslopifying your gens, right?
how about something simple to cope with artist mixing? using prompt scheduling doing something like[(@artist1:0.75), (@artist2:0.95)|(@artist1:0.95), (@artist2:0.75)]. alternating the weights between artist tags with each step
>cumfy>lora training>adapters>schedulingniggas, I just want to gen cute otoko_no_ko_\(currently trap\) draining overweight salarymen not sign up for summer break homework.
why are the gays like this
declnaine
>>8828792ngl that rose tattoo came out really well
>the pedo immediately starts boilin' and bubblin'Always the case.
multi-thread pedo podracing
>>8828904try /hdg/
I feel like "artist mixing" on anima just makes the output regress towards the slop average rather than create a new style.
I am once again asking if it'll ever be possible to train anima loras with 8 vram
>>8828939it's definitely worse than CLIP based models
>>8828948I haven't actually looked at the clusterfuck of anima and what vram usage actually looks like but the model is only 4.18GB and the TE is 1.19GB with a 256MB VAE.so I would guess maybe? probably? at least batch 1 should work? worst case it might oom by some small enough amount that it'd still be feasible to do a train overnight. No idea how unet only training on it would work in terms of function but that would also improve the chances of it working, probably by a lot.
>>8828951Yeah that was also my logic but the one anima trainer available says it won't work below 10 vram for some reasonI guess I'll just have to wait for it to be optimized
>not only 1girl, but also 1artistlmaoimagine having this much potential and throwing it all away because your text encoder is too retarded
>>8828964architecture issuenai 4.5 has the same problem and it doesn't stop it from being nsfw SOTA
>gather a dataset of your polished illu gens of your favorite artist mix(es)>train anima lora on themExplain why wouldn't it work>inb4 training on a synthetic dataset
>>8828967it would work but training on a synthetic dataset is bad
I still don't understand why there are separate categories for illustrious and noobai loras on civ when they're essentially the same thing and are backwards compatible
>>8828972because you are stupid
>>8828968And you're missing out on potentially tens of thousands of mixes, if you have to spend time generating the dataset then training for each one.I'll still do it though, got a pair of mixes I've been dragging along since early NAIv3.
This question has probably been asked a dozen of times by now, but what are the (((optimal))) settings for Anima? Anything other than er_sde hurts the artstyle, and anything other than simple DRAMATICALLY hurts prompt adherence in my experience. What about the tags? I'm currently using "masterpiece, best quality, highres, very awa, score_9, absurdres, @artist" and THE schizoprompt for negs
>>8828985i like the full monty of pos/neg but no pos/neg is also pretty good for artist replication. and i'm using the same thing as always, 3m sde with high steps (50) and it works fine. you don't need the ratings tags ever, it gets them from context, if you do 1:1 with explicit and nothing it will look basically the same and the other ratings tags either have minimal or way too big and unpredictable effects. artist tags in negs are extremely powerful so be careful.the most important negs by far are the two non-anime datasets: ye-pop (photos) and deviantart (deviantart). if you don't have these in your negs you are fucking stupid. everything else is negotiable.like, both of these are kinda correct and kinda accurate to the artists prompted, partially, but they look quite differenthttps://files.catbox.moe/4qde6a.pnghttps://files.catbox.moe/tz9avo.png
with single artists i would go for no negs, it's extremely accurate then
oh yeah, i have also noticed that even some very popular characters are inaccurate if you don't include the series tag as instructed
>>8829006That happens for some characters in noob too, like Coral from pokemon horizons (and then CLIP will add some butiful corals for you anyway).
>>8828965true, what stops nai from being nsfw sota is the ugly flux vae, the lack of pretraining and the tiny unet
>>8829008yeah but for example if you prompt just fate testarossa and no series tag you can get her without the distinctive asymmetric hair floof the character designer does
>>8829012>fate testarossahow does this not cause saibableed
>>88288404.5 can look alright, it's just that it doesn't look good by default unlike v3 and it's mostly used by completely clueless people. Artists do mix but the mixes are unstable depending on the prompt. It's still kinda better than anima mixing (although I only prompted anima for like 2 hours or so. Trying shit like prompt scheduling might be worth a try later)
>>8829006same case in chenkin noob
>>8828721>Whats the source on this?wdym source? it's ai generated
I.... neeeeeed..... anima full releas---------BEEEEEEEEEEEEEEEEEP
This but uncensored z-image danbooru edition
>>8829019>BEEEEEEEEEEEEEEEEEPmicrowave?
>>8829020Stay realistic
>>8829013she has almost 10k images dog...
what if we trained anima on real images
>promp Fate Testarossa>get Seibah driving a ferrariThank you CLIP, very cool
what if we trained anima on every single fucking furry and cosplay image we can conceivably get our hands on and added some screenshots from miside for good measure?
>>8829035now that the dust has settled, why did lax do this?
>>8829035miside is appealing
how come models improve but gens never?
>>8829035Explain how that made noob worse, providing practical examples and not just crying about the sanctity of anime being sullied by evil Western and 3DPD images.
>>8829034hilarious and original
>>8829044WHEN YOU INCLUDE NOT ANIME THINGANIME GET LESSVERY COMPLICATE
>>8829033I'll do it as soon as the full model and a training script that isn't vibe coded tash drop, I have my BLACKEDRAW dataset ready!
>>8829047I said>not just crying about the sanctity of anime being sullied by evil Western and 3DPD images
>>8829050it's not fucking spiritualit's how imagegen models work
>>8829052Ok, can you show some examples of noob being less anime than illustrious? Or 1.0 VPRED being less anime than 0.5 eps?
>>8829043New tech can't fix bad taste unfortunately
>>8829043Look at 1.5 gens and say this again.
>>8829054oh so it's pretend to be retarded time again
>>8829058I accept your concession.
Anybody here tried gay training an Anima lora yet?
>>8829063>gay training
>>8829063No, it does my favorite trap artists pretty well out of the box, albeit shuffle (songdatiankong) switches between his old and new style pretty randomly.
I hate that gay people exist and come here
>>8829067Same
>>8828985worst_quality in negsmasterpiece in poseuler A/NormalGood enough for me.
nothing gay about wanting to fuck otokonokos thoeverbeit
>>8829050The sanctity of anime being sullied by evil Western and 3DPD images is a bad thing doe? Where do you think you are?
>>8829073True, I can't argue against that as long as is cute
>>8829067this.
>>8829073eurocuck spotted
any game recommendations?
>>8829090mary skelter
>>8829090nioh 3 is pretty good
>>8829090what type of games do you like to play
>>8829094i like terraria and pokemon games, anything i relax and take my time with
>>8829090I'm playing Nioh 3 and having a blast
>>8829095i should add that i also like to masturbate
>>8829095Factorio if you aren't a brainlet
>>8829101Even if you are, just weaken/disable enemies.
>>8829095darkest dungeon is a very relaxing game and you can add hentai mods to it, it look really well once you have everything set up
>>8829095stardew valley with this mod https://www.nexusmods.com/stardewvalley/mods/29979
>>8829090genshin impact, zenless zone zero, arknights endfield
This turned out pretty good.
>>8829124would have turned out gooder is she had bigger tits
Another one
>>8829128at least remove the watermarks
>>8829123yeah i would euthanize you
unc def in need of some zen in his zone
>>8829121hot.
I thought claude was some crypto bro meme but so far I>made a chrome extension that lets me fully automate my novelAI set creation>can give it a bunch of images from a previous set I made and just swap the character prompt out>Made an extension that lets me streamline censoring and watermarking imagesI may actually be able to triple my grifting with this and just make two spinoff patreons with different styles and nobody would be the wiser.
buy an ad
Just came back after abandoning imagegen for a year to test Anima.Are Easy Scripts and Kohya GUI abandoned? Is that diffusion pipe thing the only option now? I don't want to install WSL just to train loras...
>>8829160it's just too early
>>8829181I went to check in with 5ch because I remembered that you guys are useless, and it turns out that they figured it out already:https://github.com/RedRayz/Kohya_lora_param_gui
>>8829160it's only been like a weak
>>882916067372a fork of easy scripts has anima training support
https://exhentai.org/g/3719205/695938b410/What's your excuse for not genning an entire doujin?
>>8829191Good to know, thank you.
https://files.catbox.moe/ksl3yt.pngHelp I can only make retarded shit like this instead of using the power of NL for good
>>8829213based creativity broken futachad, now make caesar fuck burnice or something like that
>>8829186I just gave this a quick test and it works pretty nicely even with 8 VRAM. Still need to spend a bunch of time fiddling with the settings, of course.
>>8828924Been here since #1 and stuck around after the decline of #10, tourist. You should go back.
>>8829044e621 tags are a godsend sometimes. There, I said it, chat.
>>8829236based farting_on_prey enjoyer
>>8829236trvke
>>8829236Trvke. Genitals also suck on Anima, just like they did on NAIv3 and Illu 0.1.
>>8829192i did it once and making all the panels is hard workid rather bake a lora
>>8829236the anime images should be rescued from e621 hell
>>8829251if you love anime so much, use animagine. it can't do porn at all, but it's pure!
Honestly, gay furry porn is superior to lame anime girls.
why are you prompting for gay furry porn
Pulchra fellini should be allowed on /h/
>>8829249>Genitals also suck on Anima, just like they did on NAIv3 and Illu 0.1.Oh fuck. I forgot all about that. I think there was even fucking penis loras going around back then.
>>8829291Enjoy your pure anime btw. Properly tagged and filtered e621 would have no negative effect whatsoever, but that would make the schizo angry regardless. People have ptsd from pony, but that piece of garbage wasn't an anime model to begin with, it was a sepia cartoon porn model.
>>8829342and anima is already contaminated with deviantart and ye-pop (whatever the fuck that it) for some reason, those two shitholes literally add nothing but poison the dataset. at least e621 makes nsfw part better.
What pose/position to do next?
>>8829372Why's that orc wearing a piece, Sir Solangeanon?
>>8829265sadly i am impure
>>8829291yeah, a penis lora that didn't even work
>>8829343ye-pop is photos, which if tagged appropriately are okay for shit like backgrounds (see sd 1.5)
>>8829382>>8829387
>>8829393i mean just throwing a few photos in is probably just gonna poison the dataset
not to mention i recommend negging it out in anima specifically, it looks like dogshit without doing that
>>8829393realism causes the grease but not sepiaalso if you completely 100% exclude realism you end up with something like nai 4.5 that requires 10 artist tags to get any sort of gradient shading
>>8829396what a fucking shitshow nai 4 and 4.5 were
>>88293974.5 is workable
>>8829396like base illustrious
>>8829396don't you worry. those score_9,score_8,score_7,score_6 tags will make sure you get gradient shading whether you want it or not
>>8829406base illustrious was pretty bad but you could still just put in a fucking artist tag and it would work
>>8829406base illustrious still had some sort of leftover sdxl knowledge and also didn't have danbooru realistic pics thoroughly removed like it happened with nai. honestly I really liked illu 0.1 aesthetic even if it was mostly only useful for 1girl.
>>8829413>have danbooru realistic pics thoroughly removedYou mean like cosplay photos? booru "realistic" pics are still just drawings
>>8829416yeah I wouldn't be surprised if nai partly filtered realistic drawings as well though, they were really paranoid about it with 4.x
If you're going to poison the dataset with /aco/ sepia garbage anyways then why not add the part that is proven to make models better at porn? I don't get it.
>>8829437this is a good argument
when are furfags going to finetune anima on their shit. finetune scripts are arleady available
>>8829481After they finish training Chroma.
https://files.catbox.moe/v622v5.png
>>8829524>only 1 is cosplayinga bit strange
>>8829533it's no metadata found
>>8829533I expanded the idea of an old gen that was originally only kuro, illya didn't look as good on a taimanin suit so I went with her magical girl outfit instead>https://files.catbox.moe/3hxbjc.png
>>8829382it's photos and other shit like game screenshots, only worth mentioning because he said he got rid of the photos
>>8829396>gradient shadingIt's easy now, with style reference......
>>8829224But plain fucking is boring.
>>8829283she even got banished from the boorus to hidden furry tags recently...
best sampler combo for anima?
>>8829623i see no reason to swtich from 3m sde sgm uniform 50 steps
>>8829624>50 stepsaugh
Thank fuck neo got updated to support anima, now I can ditch c*mfy until a new exciting new model comes out.
>>8829627you can get by with 36 but sde loves its steps
>>8829628big if true
>>8829630It's true, all of it.https://files.catbox.moe/dpjph7.png
>>8829634why are you gay
>>8829636What's gay about wanting to fuck gacha sluts?
>>8829634the power of the 16ch vae in full use trying to render that old man's face
Hold the fuck upAnima doesn't have the e621 dataset? Into the TRASH it goes.Wake me up when NoobAnima (Full) releases.
>>8829685Tell the guy to add it, before he starts the full res bake. Instead of Deviantart.
>>8829558It's good but the anal cost makes it essentially useless.
>>8829623Anything other than er_sde @ simple produce terrible results for me, but I'm 70% sure that it's a prompt-related skill issue on my part, I'm still experimenting with it>>882962750 steps with this thingie at default values https://github.com/Jasonzzt/ComfyUI-CacheDiT is faster than 30 steps without
>>8829686I might but I don't have high expectations.And in the end the we will end up using some shitmix over the base checkpoint. I happened with Illustrious and Pony.
>>8829692huh? cachedit spits out errors for me when i try to use it with anima. Also have any of you found a good upscaling solution? straight latent upscaling gives me grid like artifacts when used with er_sde or euler a
>>8829693Sure but a shitmix won't add in all the furry knowledge, just style bias.
>>8829695Multidiffusion is the one thing that removes the weird pattern altogether, but it still breaks above a certain resolution, alternatively using high CFG values reduces the artifacts.
>>8829695i wasn't able to find a good upscaler so i just gave up for now
>>8829695>straight latent upscaling gives me grid like artifactsIt's not just latent, it's any time the model does img2img at over 2 megapixel res.I use classic SD1.5 era tiled upscaling >>8826986, this way it works with 1mpix squares one at a time.Refreshed catbox https://litter.catbox.moe/blx4e9ui79wgdmyz.pngWould be a lot easier sharing knowledge if we were all in on thread.Multidiffusion >>8829725 is a newer version of the same thing, does it work now? Last time I checked it was not yet updated for the new Qwen stuff that Anima uses, and would just crash.
>>8829700I'm not looking for furry stuff but more nsfw knowledge. Like the jump from illu -> noob, it was more than fine in that regard. And the noob shitmixes were the final touch.I wish for something like that.
>>8829736My bad, it was tiled diffusion not multidiffusion.
>>8829695>grid like artifactswelcome to dit patching baby
>>8829634What's the verdict, non cumfy bro? Anything worth jumping off my sekret 291 mix yet?
Why are we on page 6 wtf
>>8829823The usual. Being a split thread steals roughly half the activity, and then being slow it also dissuades others who might have joined in.
>/hdg/ Hentai Diffusion General #777.20w14∞What causes this?
severe lack of BBC in bussy
>>8829778It makes generating more involved character interactions significantly easier than trying to tardwrangle SDXL and CLIP for sure, and if you don't like the style it produces or want to use your sikret special artist mix you can just use the anima gen as a CN for Noob. The problem is that genitals are not as good as noob due to no e621 and it's missing those extra specific tags too which aren't easy to replicate with NL.
>>8829935well, there's no substitute for good tagging
>>8829613how do I get a job at the dildo testing facility
>>8829935I was just fooling around with it and the speech bubble, the juices and the visible pussy lips were purely from the description, no tags at all. That said it also seems to be quite literal, if you don't tag or describe a background it won't generate one, it's also much more susceptible to shitty tags
>>8829946Very nice gen.
>>8829935mixing artists doesn't really seem to be all that difficult. You can just slap in @ARTIST, @ARTIST and it'll do something, and do things different if you append them with strengths. It's a lot more noticeable with vastly different artists' styles of course. it's definitely a lot fucking cleaner with what it gens though, it definitely does a better job of masking that kinda smudgy vectorized look of other models, kinda shocking that this is a smallish preview model
Is it too late to add e621 and shit to improve the model's NSFW performance? Is the preview a snapshot of the full model from mid-training or more of a proof of concept and rather than be a direct continuation, the full model will be trained from scratch, potentially with an expanded dataset?
>>8829955you should ask the model creator
>>8829954I meant that if you type in the exact same mix it won't look the same and it's probably going to be impossible to replicate.
>>8829960obviously...it's a different model
>>8829955I'm wondering the same about the TE, 0.6B is just comically small and would be a massive mistake to go ahead with for the final model.
>>8829955define poor NSFW performance, out of the box it already reproduces various prompts/concepts that Illustrious 2.0 had no idea what to do with
>>8829964ntathe genitals are not that detailed and prompts at nonstandard camera angles frequently cause misaligned genitalsalso illu 2.0 is dogshit
so glad im subbed to nai
Everyone who's whining about anima being terrible are just wrong. What people don't get is that the quality of hands, anatomy, genitals or x is highly dependent on what artist/style you're using. As a monster girl genner this model is miles better than the shit I could gen with Noob/Illu.https://litter.catbox.moe/st4rl1q0fd7b9mfx.png
>>8829968very funny, but actually i will note that closeups in general are very good on anima. and oral in general. illu has some serious problems differentiating a deepthroat and something else but this model doesn't.
also nobody's saying anima is terrible, the sentiment is generally very positive. and everybody knows it's a 512x version
>>8829968>Everyone who's whining about anima being terrible are just wrongI haven't seen anyone whining that it's terrible doe?
>>8829968>"anima is very good for a preview but there's definite room for improvement here and here" = "whining about anima being terrible"
>>8829972how leftoids see the world
How hard would it be for cumrag to implement a readable prompt info string at the beginning of the metadata as a default behavior? Fucking retarded faggot
>>8829974Forge Classic added Anima support, there's no need to continue to put up with cumrag anymore.
>>8829975i think all the forges got it now
>>8829976reForge was last updated 3 weeks ago.
>>8829977okay not abandoned ones
Forge Neo or Forge classic? I've been on a March 2025 reforge commit for quite a while.
>>8829979i've been using neo and the removal of a lot of garbage is good but they also cut the ability to save controlnet presets which is really useful
>>8829974make a node
>>8829965yeah the misalign thing does happen but it's not that bad, as for detailed genitals I dunno, cocks look like cocks whether in gens them censored or not, gaping vaginas are kinda silly looking as it doesn't seem to render cervixes well but it does well enough for your standard penetration. Tentacles decidedly generate penetrating inward as opposed to pushing outward from orifices like I used to see a lot of, although it does need some prodding to generate monster tentacles over squid tentacles
>>8829989i've found it's somewhat tough to get tentacles in to mouth as opposed to out
>>8829955you won't get an actual answer unless you ask here https://huggingface.co/circlestone-labs/Anima/discussions
>>8829993
>>8829966>>8829997stay at /hdg/, mkay?
pay me 25 dollars
>>8829999nice trips but sorry my 25 goes to my oshi every month for that tier 3
those are quads you blind localkek
>>8830002sorry i just woke upi will give you one apologen
Anima was not trained on e621 but can generate better and more realistic pigs and horses than noob, what gives?
>>8830008yeah but what about their cocks?
>>8830008>realistic pigs and horsesWhy would you want this
>>8829988There are nodes for this already. Issue is that it's not the default behavior and you always get this garbled json mess when you try to look at comfy meta.
>>8830008ye-pop probably
>>8830010How else am I supposed to enjoy my besitality if the animal looks like a Disney OC?
>>8830015>>8829998
>>8830017i still have no idea who that is and why you're so obsessed with him
>>8830019it's an avatarfag who used to rely on deadnoodles' styleyou can still see it in the faces sort of
>>8830037How is he an avatarfag?
>>8830039You're right, I was trying to avoid using "pillar" but ended up using the wrong term. An identifiable poster, but though a consistent style not through a character.
Quick browse through the thread (hey, actual discussion for once) so if I read everything correct>every Forge but Pancho's now supports Anima?>base anima gens can only kinda somewhat be upscaled through boomer SD Upscale?
>>8830037>>8830041doesn't explain the obsession
>>8830037Well, you are making him seem like a pretty decent chap.
>>8830048when the actual model comes out someone will train a tile upscaling CN and it will be all good
>>8830051he is, people are just funposting
>>8830052Here's hoping. Guess I'll give it a try over the weekend so I can get familiar with the prompting style.
>>8830051Other than the pedophilia, and baking new /hgg/ threads when we finally had an opportunity to merge back, yeah I suppose to.>>8830048>kinda somewhatWorks fine as long as you keep denoise low and run the extra tiles for seam fix. Just takes a long time.
When you pay me 25 dollars
>>8830071I've been here since #1in both since then, cause you refuse to put them back together
Why is the real pedo calling ME a pedo, chatI am lost on my own lore
>>8830096What do I have to do with that dogshit place?
>>88300922bh I mixed you up with oekakianon because of the thick outlines and tendency towards young girls
>>8830099oekakifag posts toddlers with gigantic heads, noodlebro posts slender girls with tits. how is it possible to confuse them?
>>8829968wat happen to her mouth
>>8830102I see a non-hag with a thick outline I don't give her a second look
>>8830107I also post hags wtf
>>8830110The style makes her look way younger
why would someone purposefully want to look at old women
>>8830112So is not about the content but the style? lol
>>8830121Yeah let's not get into that again
>>8830123That's the problem, chat You could be posting cute gens along with me but instead you keep bringing nonsense to the table.
>>8830104She lost it as a zombie, as they tend to lose random parts of their body.anyone still doing animated stuff around here or is it only shitposters?https://litter.catbox.moe/9hzvh2h616gxhk94.webm
>>8830126Last I tried Wan video it cooked my GPU for 20 minutes then produced a still image.
>>8830099does this look young to you
there he is
here is your 1girl missionary pov slop sirs please calm down
>>8830145can you generate this again but without the overwhelming pissfilter
>>8830146no warm colors are necessary for a warm mood
Did the anima guys say anywhere if they are planning to train CNs for their model?
>>8830182I hope so, but given how much easier it is to control the composition now, what is CN going to add?
>>8830182you should ask on their discussion page https://huggingface.co/circlestone-labs/Anima/discussions>>8830185tiled upscalingIPAscribbleetc.tone os uses
>>8830185Mainly for upscaling, I don't really use CN for much else.
>>8830185tile, scribble, IPA and inpaint would be nice
>>8830186>>8830189IP-Adapter is not controlnet
>>8830191technically true
where can i get the optional t5 stuff? will it make training better?
>first version of noobai came out a week after the illustrious leak>all anima gets are slopped shitmergesIt's ogre isn't it?
does the tiled vae decoding work with anima?
the general rule with anime isif you have to ask, it cant do it
>>8829623i like dpm++ 2m sde with simple/sgm uniform, 4 cfg
>>8830230why not wait for the full model?
>>8830392we have no ETA for full modelcould be a year from now
>>8830121Sir you posted this image in this general before.
>>8830481fair enough
>>8830367that makes no sense whatsoever
>>8830490Yes I did
so is anima worth my time yet or is it still wait for fixes
>>8830519It takes like five minutes to get going. Quit being a little bitch and try it yourself
>>8830530That's five minutes anon could be staring at porn with his mouth agape.
Hmm.. I should gen some gaping otoko's..
>>8830558proof?
>>8830519fixes for what
>>8830569fixes for not being able to prompt farting on prey.(you probably can still, but where is e621)
>>8830570Ask Lax to get another research grant to train porn models.
It's weird how on anima some characters will not register unless you add some specifier to their appearance. Like kanzaki_ranko on its own gives me a nondescript girl but if I add even two or three related tags she gens her spot on.
>>8830587hopefully it's just undertrained
>>8830606no
>>8830607yea im retarded. i saw their csp time lapse. the style just looks like ikuchan kaoru so i assumed it was AI.
>>8830308why bother on a preview
>>8830587yeah the character knowledge is kinda bad. it's can't even properly do maria cadenzvana eve, when tagged with the series tag, and she's got a very specific hair and hair ornament shape.
and 1300 images
so what now?how can i wait until anima full?
>>8830619well...you could jack off
>>8830619Gen some POV handholding netorase with your favourite gacha slut.
>>8830619nai sub
>>8830625Hmm not a bad idea
https://civitai.com/models/2385278/animayume?modelVersionId=2682208so it has begun
Anyone got updated danbooru tags for anima? The cutoff date is much later than other models.
>>8830714At least they all have to finetune now since there's nothing to merge it with.
>>8830718>train your own lora>merge it in>???>"It'S a TrAiNeD cHeCkPoInT"
>>8830714let's heckin goooo
>>8830714yume slop slop
>>8830568I wouldn't lie to you, anon.https://files.catbox.moe/lwq8q1.png
>>8830714seems okay i guessi'll do another with a single artist so it's easier to see but the artist mixing looks okay on both
Chat, should I buy muse dash with the dlc or pump it up rise?I have played both before and I like both
>>8830736never heard of eitherblessed ixy gen tho
gonna be real this barely does shitif i had to rate it no qualityslop animayume is slightly better
>>8830743They are working over something that barely works to begin with, what do you expect
>>8830746i mean...i think it's great for a preview. by virtue of the trainer not being a fucking retard, mostly
i'll even take a wai-type shitmix at this pointjust give me something with a consistent style that doesn't break hands and anatomy please
>>8830753>just give me something with a consistent style that doesn't break hands and anatomy pleasethat's been solve since ages ago tf
>>8830755i'm talking about anima duh
>>8830753just prompt a good artist
>>8830756Oh alright yeah, better wait for the full release
>>8830757yes and then i get the right style on 1/10 seeds and prompts with a model that is twice as slow
>>8830759SHOUGANAI WA NEEEE
>>88307562 more weeks until full release
>>8830743wtf is qualityslop
>>8830778masterpiece, best quality, absurdres, score_9 and so on
How to do tiled high res in forge neo? multi diffusion is only available in the img2img tab
>>8830803comfy here but iirc it was called Ultimate SD upscale in old A1111
>>8830803You mean Ultimate SD Upscale? It's in the Scripts menu at the bottom.
>>8830778score_9 score_8 score_7 masterpiece best quality newest recentworst quality low quality score_1 score_2 score_3 old early blurry jpeg artifacts sepiatesting these now since i am no longer forced to use a DOGSHIT ui without convenient x/y plottingyou should always neg out ye-pop deviantart, that's easy to see
>>8830646>artist:firedotincWow, novelai can do anything now.
https://files.catbox.moe/gzyljp.png
>>8830806That only shows up for me in img2img as well
>>8830814more ffm threesomes
>do one x/y plot>score 8 best one again same as fucking pony>equivalent in other tags (best quality) also fineamazing woaow who could possibly have predicted this
>>8830817I just noticed that, I have no idea then.
>>8830806what settings to use for it?
>>8830783>>8830807I only use>best quality, masterpiecein that orderbest quality because it had minimal impact on style but gave more consistent anatomy, and masterpiece after because otherwise it would overpower composition and style but going after best quality it only adds a bit of flair don't use score tags, they're fucked and will give you shit shading and just make everything more 3D looking
>>8830838there's no practical difference between score tags and the NL quality tags, they are associated with the same images.just did an x/y on negs and some of them totally rape the quality while the ones you'd expect (score 1 and old) are goodhttps://litter.catbox.moe/vxud24atlpn0vnbx.jpghttps://litter.catbox.moe/j43erhb686v91712.jpg
oh and ratings tags either rape your image quality or do literally nothing. that was easy to test
>>8830842>there's no practical difference between score tags and the NL quality tags, they are associated with the same imagesUnlikely to be true, masterpiece and best quality and so on are based on the image's score on danbooru while the pony score tags used astra's aesthetics raterIf you want to see the real difference between them don't prompt any artists.
>>8830847i will test that later then, rerunning on the base model first
>>8830851Here's a really simple test, all images include highres and absurdresno quality tags, masterpiece, best quality, score_7, score_8, score_9, score_9, score_8, score_7
>>8830857not a fan of highres/absurdres they are too unpredictable
also>testing on only one seeddo not reply to me further retard
This is the kind of sovl uncond you get when your dataset isn't poisoned by e621.This is what furfags have kept from you since the days of pony
>>8830859(you)
>>8830861and what is noob uncond? horse smegma?
>>8830861you have no idea what you are talking about
>>8830847Did he ever clarify that? All it says is >PonyV7 aesthetic model based: score_9, score_8, ..., score_1Which could mean the tag names are based on pony, not the data tagged with them
>>8830842I'm surprised how similar they all are, can't even tell the difference. Maybe not an ideal prompt for the test.
quality tags are bloat. just use loras and artist tags and that's it
>>8830874In context with what he said about the standard quality tags I think it's fairly clear what it means
>>8830842dont do them in negs, try them in your positivethere is a major difference between score 9 and masterpiece for example
>>8830857yeah as I said, score tags give you that 2.D pseudo 3D AOM slopdon't use them
>>8830889>2.D2.5D
>>8830876if you cant see it in a real prompt then waht's the fucking point
>>8830879in general i agree
>>8830887why would i test negs in a positive promptagain if it doesnt work in practice then what is the point
>>8830900if you can't see any difference then it's either a bad test or there was no point in testing itat least in hindsight
>>8830920this is a non sequitur, you can clearly see minor differences especially in the crispness of lines and detail level of extremities
it is what it ishttps://litter.catbox.moe/p3zvyoa63gsyn16b.jpghttps://litter.catbox.moe/p3zvyoa63gsyn16b.jpg
well it turns out anon was completely right about score tags showing up differently without artist tags https://litter.catbox.moe/k7i2w6kw7gspb9in.jpg
>>8830934they always just carry style bias of their own, the models don't learn the abstract concept of "quality" but just and averaged style of high quality images
>>8830912because it lets you know what is actually associated with the tag?that way you'll know if you even need to neg it or not
>>8830934>the ugly inherent style of score_9, 8 or 7ponybros are eating good
>>8830937that wasn't really my point, you don't need to explain the basics to me, this is my point >>8830940
>>8830939well fair's fair and idk even how to feel about this one https://litter.catbox.moe/zkyn4j26h2i51cas.jpg
>>883071430k steps
>>8831020giving the steps means nothing without also giving the batch sizeis he stupid?
>>8831020470k steps to go
so tldr on quality tags/negs?what's best combo on minimal style influence but good anatomy/interesting composition?
>>8831025best quality and masterpiece both have a very high impact on anatomy quality and coherence imo, higher than on almost any other model i've usedi usually like to stop using quality tags once i've trained a lora i like but i'm having trouble doing that with anima
>>8830587>she gens her spot onanima is a girl? god I need to know what she looks like
>>8831025masterpieceye-pop deviantart, worst qualitythat's itDO NOT use high score tags, they are the shiny ponyslop look
random cumplot from skmming boorusome of these look pretty terribleforgot i had an artist neg in there tho SHOUGANAI WA NEEEE
>>8831175forgot my image on that shit https://litter.catbox.moe/7ccnvhfkqf3hz9dj.jpg
also im deletin animayume shits weird lol
We need 291hanon to do for anima what he did for noob.
>>8831181Merged a bunch of random checkpoints together? >>8830718
>>8831181trueeee
falseeee
proofs?
We need autismanon to do for anima what he did for pony.
>>8831252is there a lora of this style yet
>>8831255>https://civitai.com/models/723360/ai-styles-dump-animaillustriousrouweinoob?modelVersionId=2466040
>>8831259>all glued togetheri'll just have to do it myself then
training anima on ezscripts doesn't work
What can anima do that noob can't? This looks more complex to train loras with considering the confusing "natural prompt" thing, it just feels dry than direct tags.Also any gui for anima training yet?
>chat would rather train a lora on my images (again) than create a lora of chinatsu or velveti am so done with life
>>8831265shocker
>>8831266doesn't have eps grey or vpred frydataset is updated to early 2025prompt comprehension is good due to qwen vaeartists are accuratewhat's not to like?only problem is it's a 512x preview version so extremities can be kinda bad, and it's undertrained especially for characters
>>8831267give me the dataset and i'll bake itstanding offer to anyone who makes and tags a datasetunless it's for something cringe like piss
>>8831266Anything that's more complex than "1girl, 1boy". You have so much more control over composition, positioning and actions with NL. Gens like >>8826799, >>8829603 or >>8830814 would never in a billion years be possible on old architectures outside of unicorn seeds.
>>8831268I had expected it to work after the guy added support, but I'm either to stupid to check the right boxes or it's still buggy
>>8831269I want to import my lora training images to anima and see if its superior, however the waiting game for a gui might be long. Hopefully my little 16 vram is ok.
>>8831273there's a reason i gave up on ezscripts
>>8831272>>8831275Yes but don't we have regional prompt for this? It works pretty well imo.
>>8831277Regional prompting by itself is a cope that's janky to use and simply doesn't work more often than it does. If you want it to be useful, you NEED either NAI's macros (target#, source#, mutual#) or NL, otherwise you're still at RNG's mercy for the actual interactions.
>>8831278brainlet who doesn't understand the syntax detected
>>8831277you don't need region prompt with a good vae. literally just type in "the girl on the left xxxxx" or whatever the fuck
>>8831281>vaeYou mean UNet and text encoder. VAE is responsible for upscaling the already generated latent space to full res so a better VAE means better details, but it has no effect on the overall composition of the image, which depends on both the UNet (visual representations) and the text encoder (textual representations).
>>8831286okay
>>8831277regional prompting is awful
regional prompting fucking sucks dick and cope.loras fucking suck too.
Still better than anything $25 can buy tho
>>8831296>loras fucking suck too.TRVKE NVKEMSuch a pain in the ass to make, and you can forget about literal whos with only their one official art.
>>8831297FACT CHECKED: TRUE
https://github.com/kohya-ss/sd-scripts/blob/main/docs/anima_train_network.md
>>8831329lets go
How good is Anima's natural prompt comprehension? Is prompt bleeding finally solved because now you can just specify "the girl on the left has blonde hair, big breasts and wears a red micro bikini while the girl on the right has red hair, flat chest and wears a blue slingshot swimsuit"?
>>8831366it is fairly reliableuse more NL if you want more accuracy
As there's this discussion with regional prompting and poses and scenes, I'm having fun with the Comfyui Panels node, though it's a bet on getting good enough gacha rolls 4 times in a row. Works well enough with simple scenes, but add 2 very different girls for example and it becomes nightmarish to keep characters consistency across scenes. I'm using Illustrious.The original example uses Qwen and nodes that seemingly allow to keep that consistency across scenes by feeding it base images of characters, I don't know much about Qwen other than it being xbox hueg, I think it's still not SDXL-anime-booru-friendly, right?Upon searching on how to keep consistency with those models, it seems I'm, out of luck. Controlnet/IPAdapter/somethingIDface or whatever exist, but I feel like I'll be chasing a chimera.Is consistency the Aimeme endgame? Any tips on achieving that without it being overly complicated?
>>8831265are you using machina's branch? does your gpu support bf16? then there's no reason it shouldn't workwhat error are you getting?
>>8831366Just genned a batch of 4 pics, no cherrypicking. Catbox because not /h/, also I'm too lazy to combine themhttps://files.catbox.moe/ye7j5s.pnghttps://files.catbox.moe/5uu6w3.pnghttps://files.catbox.moe/ovp2xm.pnghttps://files.catbox.moe/ere8al.png
>>8831370you could do that with regional prompter 10 years ago. gen some entangled group pose and make the outfits more complex
>>8831371>10
>>8831371I can't believe sd1.5 came out 20 years ago bros...
somehow i get the feeling artist mixing on anima you need to reverse the order
>>8831369https://pastes.io/error-mode
>>8831377idk your install might be fuckedmaybe try to install it again in a different folder
Multidiffusion works for anima on neo now btw
>>8831379can you drop me a catbox or settings page or whatever
>>8831379still only in img2img?
>>8831382Yeah.>>8831381The default settings seem to work fine.
>>8831329kinda confusing, how would you make your own toml files? I also want a unet only type of learning too via Glora.
any anima config tomls yet? I don't want to start from scratch
>>8831400i've just been using a very barebones adamW 1e-4 constant and it mostly worksbut experimenting in general is hard since it takes twice as long
>>8831398the toml is just dataset config stuff, there's a link to a page explaining that too. i kinda hate having two separate files for dataset configs, i think it's retarded
>>8831415yeah i'm trying to figure this out, the gui is currently never ever and I want to try porting my loras to anima for testing. all these .py and scripts are really confusing, even one of the docs trying to explain it to you is empty via train_network.py guide -> dataset preparation. literally file not found in the docs. guess i'll wait for more guides or someone passing a full config for this .py thing.
>>8831419it's actually quite shrimple you just make a text file with the extension .ps1 and then put all the arguments and shit in the text file, then open it in powershell and run it. the toml is just another text file same shit
>>8831419>the gui is currently nevermachina fork of ezscripts has a gui and can train anima
>>8831426ok i'll bite
>>8831426>go to repo>check forks page>ctrl-f machina>nothing shows upsasuga /hdg/
>>8831434https://github.com/67372a/LoRA_Easy_Training_Scripts/tree/refreshanon I....
>>8831436that says 67372a not machina though?
>>8831438dude goes by machina on trooncord. didn't expect a complete newfriend
>>8831440i've been here since the beginning but i have a mental block on discord trannies
>>8831448what causes this
PANCHOVIX WHERE'S THE ANIMA SUPPORT FOR REFORGE REEEEEEE
>>8830008it's the devianart masterpieces.
>>8831451Nigger what are you doing still using that abandonware
>>8831454Last commit was yesterday....
>>8831454fuck off noodle shill
bakaschizoadd cccino to your list
>>8831457I think he means you should be using forge neo, which does have support>t. comfy user
p-panchito.. *cough* o-onegai.. just a drop of.. *cough* a drop of anima.. k-kudasai, panchito.
>>8831539literally no reason to use reforge over neo
>>8831539give my negro some time, I am sure he will do it maybe by the time the full model releases
>>8830857wow best quality is also slopped. grim
>>8831579I wrote that post really poorly, none of the images is just best quality on its own, the 2nd image was masterpiece and best quality together
>>8831586oh i see. all good
>>8831451>>8831539he won't do it. it's massive amount of work to rewrite the backend to support LLM te, ect. just use neo if you're allergic to cancer noodles
>>8831597But neo gives me worst gens than reforge
>>8831603does reforge give you an anima gen?
anyone done extensive testing on what negs are good?
Panchobros.. I redeemed the neo. >>8829213Thanks for the metadata, anon. No clue what I'm doing yet but your shit at least put me in good starting point.>>8831613Scroll up. I'm just going with what anon >>8831163 suggested for now. At least until I get my feet wet.
>>8831609I'm not joking btwWhy the fuck neo gives me worst gens than reforge? I would have made the switch sooner if it wasn't for that
>>8831614i find year 2025 to help lock in newer styles for some artists.
>>8831616who cares. there are 0 reasons to use something other than anima now
what does anima bring to my 1girl standing workflow?
anyone else not getting nan loss when training anima with ezscripts?
>>8831635She can hold up a sign with "BBC ONLY" written on it
>>8831635better prompt comprehension and colorsearly 2025 datasetsome characters just don't work tho even with over 1k images, i think it's undertrained
>>8831706It's not just "undertrained", it's a preview that's not even halfway through its training. The fact that it can handle SDXL resolutions while being trained almost exclusively on 512px res pics is a miracle
>>8831712which is why people are excited
>>8831714>people are excitedLike who?
>>8831715literally every sentiment on anima is positive if you actually read the threads and comment sections
not a single person is saying it's dogshit unusable trash, even the people who don't like it that much aren't shitting on it
>>8831716Has CatHag shared her opinion on it yet?
pls respond
Someone tell lax to start training anima on the farting_on_prey and bristol_type_3 dataset already.
Only exciting thing about anima is the non flux 16ch vae. natural language and newer dataset are convient imo, but you could achieve the same with loras and controlnet
>>8831723it better have all dataset images reencoded to lossy webp again or no deal
>>8831726>natural language and newer dataset are convient imo, but you could achieve the same with loras and controlnetYeah but it's great qol and reduces the times you have to reroll to get certain specific things that aren't well covered by booru tags.
>>8831726>loras and controlnetkill your fucking self
>>8831730does the NL activate to compensate if you try to write a tag that isn't properly known?
>hating on loras in the baking thread
>i don't mind my models lobotomized so i can apply 30 patches i made myself and fry the fuck out of the model
>who are we quoting bros
>>8831732nope
>>8831732I don't think there is a functional distinction between tags and NL, it's all text to be tokenized.
>fryskill issue>needing more than one loraskill issue
why are loratards like this
>>8831740It's actually hilarious that they don't understand that nobody is hating on loras as a tool, it's the freetards who say "use controlnet and loras lol!!!!" as a response to everything that are being mocked.
are we really gonna do this same shit we've done since november 2022
any thread where anything remotely good is said about local irks nai shills to a comedic degree and they have to shit it up.
way to prove my point
>>8831755>>8831754>>8831753>>8831752can you refrain from this lowtest passive aggressive posting and properly discuss shit with a reply chain? that way i can hide it properly.
>>8831754you forgot to call them goyim animals that deserve death
>>8831763I thought n*i where turks, not jews
i tried to do a inset with gendou's face in it but surprisingly noob cant really do his face
>>8831770Based evaschizo. Still not trying out anima?
>>8831771im still trying it out, in fact ive made several posts in this thread asking about samplers and such
>>8831772Any findings? I've been jumping between ER SDE and DPM++ 2M. Both with SGM Uniform. I seemed to have less finger/toes issues with Euler A++ but it does indeed lean a bit towards that 2.5d look. Not overwhelming but notable. Of course I'm on Neo so you probably have a myriad of more options on comfy. Sampler options on Neo are limited compared to Reforge.
>>8831776>Sampler options on Neo are limited compared to Reforge.you can just add the extra sampler extensions back if you want.
>>8831763I mean, they still don't understand why they are being ridiculed. Nobody hates local (why would anyone hate free shit especially if it's good?), everybody despises dumb freetard cattle.
>he's still mad
>>8831777Where would a retard find this, anon? Had enough of an adventure getting Neo running as it required a shit ton of backend updates due to all the shit it supports.
>>8831781I remember this one https://github.com/DenOfEquity/webUI_ExtraSchedulers/tree/neg and this https://github.com/Koishi-Star/Euler-Smea-Dyn-Sampler and this https://github.com/MisterChief95/sd-forge-extra-samplers seems to be something new for forge neo.
>>8831694you are probably using a gpu without bf16 supportin particular you can't use free colab to train anima
>>8831802nah, I'm using a 3090. loss metric works fine when training with sd-scripts
>>8831803i don't know then, works on my machine
>>8831694enable full bf16 training
>>8831694I had that when I forgot to change mixed precision from fp16 to bf16.
>>8831694i tried to install this shit yesterday but accelerate refused to install correctly, it's all so tiresome
>>8831868hyouuma in this mix?
>>8831872No mix, just pure hyouuma.
>>8831855Have you tried asking claude? these installer errors are usually easy to fix and down to python/cuda version mismatches or outdated requirements.txt
>>8831868>>8831872I tried throwing in hyouuma into a mix myself but it overpowered all the other artists easily. So far, the prompt scheduling niggas are on the right track with anima. It's the only consistent way I've gotten mixes to 'work'. It's no noob but at least it shows something is happening and it's not just 1 artist overpowering the other 2/3.
my first lora with the kohya defaults turned out so good it's scary
Post it if it's otoko_no_ko friendly.
>>8831947what are the kohya defaults? can you post the toml?
>>8831966accelerate launch --num_cpu_threads_per_process 1 anima_train_network.py \ --pretrained_model_name_or_path="<path to Anima DiT model>" \ --qwen3="<path to Qwen3-0.6B model or directory>" \ --vae="<path to Qwen-Image VAE model>" \ --dataset_config="my_anima_dataset_config.toml" \ --output_dir="<output directory>" \ --output_name="my_anima_lora" \ --save_model_as=safetensors \ --network_module=networks.lora_anima \ --network_dim=8 \ --learning_rate=1e-4 \ --optimizer_type="AdamW8bit" \ --lr_scheduler="constant" \ --timestep_sampling="sigmoid" \ --discrete_flow_shift=1.0 \ --max_train_epochs=10 \ --save_every_n_epochs=1 \ --mixed_precision="bf16" \ --gradient_checkpointing \ --cache_latents \ --cache_text_encoder_outputs \ --vae_chunk_size=64 \ --vae_disable_cachethe only thing i changed was disabling cached te outputs so shuffling works
>>8831967thanks
>dumb webui-tagcomplete nigger dev hasn't updated the csv in 2 yearsREEEEEEEE
>>8831977you can just look for a newer one? https://civitai.com/models/2018479/danbooru-tag-complete-csv-tagcompleteand
>>8831977>>8832008I refuse to tag it as "trap", chat.
I've seen conflicting information on setting discrete_flow_shift=1.0 vs discrete_flow_shift=3.0. Is there anybody who has tested both and has a clear preference? I should also test that myself and compare but anima loras take too long...
>>8832023larger shift emphasizes noisier timesteps, meaning the model has relatively more time to learn medium-large featuresshift=1.0 is technically sort of how pony/illustrious/noob/any other ddpm models were trained
>>8832028That's useful, thanks.
>>8831914i'm like 99% sure it was wrong python version i'll figure it out tomorrow
>>8831967can you post your dataset toml
>>8832023After going ahead and doing the test it does seem like discrete_flow_shift=3.0 was the reason for a lot of weirdness I was getting with my anima loras. No idea why anybody was recommending using that.
>>8832120proofs?
CatTowerbros we're SO fucking backhttps://civitai.com/models/2383017/anima-cat-tower?modelVersionId=2688353
>>8832124>this explains why CatHag has been so quietShe's been cooking, bros.
>The composition utilizes the golden ratio to position the figure against the vast urban sunset, creating a powerful silhouette that speaks to ambition and reflection. Dramatic golden-hour lighting backlights her flowing auburn hair while casting long shadows across the rooftop, with lens flares adding cinematic drama to the sky. Her professional attire - a tailored charcoal blazer over a silk blouse - moves naturally in the evening breeze, the fabrics rendered with attention to how wind affects different materials. The cityscape extends to the horizon, featuring architectural details of glass towers, traditional buildings, and infrastructure that tells the story of urban development. The artistic approach combines architectural photography principles with character-focused narrative illustrationbros.. we've been outskilled.
>>8832124trained exclusively on the sloppiest images known to man i see
>>8832124omg it's so butiful sirs
>naisissoes already sizzlin' n' bubblin'>it's only a preview modeltick tock.
>>8832124>Anima Cat Tower is Anima's fine-tuned model.>fine-tunedX to doubt
>>8832124holy sloppa
>kneepit chads will never go hungry againbros.. the mad scientist autism tagging I'd have to do to even get the penis in the vicinity of the kneepit. Now I just ask it and it gives me the full squeeze and pump experience.https://files.catbox.moe/58en5m.png
>>8832170Why the fag shit?
chat, is this bait?
Anima is probably the first anime model where schizoneg will actually be beneficial, since it's not a regular clip with limited length capacity
>>8829974already exists
>>8832174Any common usage you'd recommend or do you mean in the long run with the full model?
>>8832176Not yet, I'm just gonna try dumping the whole og prompt from like late 2022 and give it a try later
go to the bathroom
why
>>8832128lmaooo
>>8832128I remember thinking that prompting like this was okay when flux came out
>>8832054sure, but there's really not much to it.[general]resolution = [1024, 1024]enable_bucket = truemin_bucket_reso = 512max_bucket_reso = 2048bucket_reso_steps = 64[[datasets]] [[datasets.subsets]] caption_extension = ".txt" image_dir = "" num_repeats = 1 shuffle_caption = true keep_tokens = 1
>>8832240oh so it really is the same as before okay thanks
WE'VE NEVER BEEN SO BACKhttps://civitai.com/models/2393785/ntrmix-or-style
>anima full releaeETA never
>>8832244Thanks, I just wanted to test how loras at negative weight work with Anima
>>8832244Now we needa WAI style lora and anima can finally be called a real model
>>8832244>>8832248>>8832250>If the creator of NTRMix @tkio wants this taken down then I will. You tourists DO realize tkio came from the general before the split.. right?
>>8832262pipe down unc :skull:
>>8832262nobody asked
>>8832244fucking disgusting piece of shit>>8832248based
I NEED the tile CN for anima, I'm no longer asking
So Anima is pretty damn impressive even in its current state. Was curious if it could do artists like Fugtrup or even Nyl, as the model states it can't do 3d well. Well, just tagging them alone you're not going to get far. But load up the quality slop tags, especially those score_'s, a '3d' here, a lack of ye-yop' in negs there, and you get a very impressive fugtrup and a very close Nyl. The latter is the one I'm more impressed by as noob (base, not the shiny shitmixes) struggles with those. It can do fugtrup's 2d art but not the style used in animations. Full model with a suite of cn upscaling methods should be nuts.
>>88322812x ultimate sd upscale works fine as long as you don't go above 0.4 denoise. Also "out of frame" in the positives somewhat helps the model understand that the processed tile does not contain everything described in the prompt
>>8832282>prompt Yunsang on NAI>bang-on his style>prompt Yunsang on Anima>generic flat slopI hope it's just the 512p training not giving enough detail.
>check out yunsang on booru>looks nothing like either of those images
>>8832294Artists with less than 400 pics on danbooru are very hit or miss and, honestly, I don't think additional training will improve much beyond that. Model can only learn that much with that few parameters. I just hope it'll get better at understanding concepts, poses and angles, you can always train a style lora later.
>>8832295this desublud is smoking something
>>88322850.4 denoise 2x tile upscale is basically asking for hallucinations, mismatches (like different eyes if they got to different tiles or different nail polish) or tile seams to appear.
>>8832300>Model can only learn that much with that few parametersIt's 2b parameters compared to noob's 2.6. Nai should be around the same size if you believe kurumuz claims of it being smaller than sdxl. Also a smaller model doesn't translate to there being some arbitrary learning cutoff.
who
>>8832308Just upscale with Flux Klein, around 2MP should work fine. Then do a Seedvr upscale if you need more
>>8832319post a gen using that workflow
So should I add images with tagged speech bubbles to my lora training data now? my one test bake lora degraded the textgen capabilities quite a bit (not that they were great before)
>>8832320There's like 6 million workflows on civitai
>>8832322do a gen using that workflow and post it here
>>8832319Does Flux Klein know how anime girl's vagina looks like?
unc cooked blud with facts and logic
>>8832294You're doing something wrong if you only gt generic flat shadingit knows yunsang, just that since it's trained 512x most detail is lost
>>8832329>You're doing something wrongIf noticed that the model is quite finicky about samplers, negs and quality tags. you have to give it a very specific baseline for it to even attempt to get a style right. just using the artist tag and cranking up the weight gets you nowhere
>>8832329>>8832331What are you using? For me it's sa solver + sgm uniform, cfg 4masterpiece, best quality, highres, absurdress in positive, no score_9slop or aesthetics shit, but score_1, score_2, score_3 in negs does look somewhat nicer
>>8832334>masterpiece, best quality, highres, absurdress in positiveslopped
>>8832331>>8832334I just use good old trusty Euler A, Normal, 25 steps, 4 CFGand>best quality, masterpiecein that order, but you can opt out of masterpiece, I just use it for some extra flavorand make sure to use @s, obviously or else the artist tag just wont work at all
>>8832334ER SDE with Beta @ 5,5 cfg. I'm trying to limit positive quality tags to best quality and mb some year tags. neg is more extensive:ye-pop, deviantart, shiny skin, old, worst quality, monochrome, greyscale
Never forget the fire immunity potion...https://litter.catbox.moe/exa9qm1pgil1hx8r.png
>>8832348No negs? Not even ye-pop and deviantart?
>>8832350>>8832367What the fuck is ye-pop?
>>8832369Some 3DPD dataset they used during training.
>>8832367I forgot what I have so I didn't mention them, I can check
>>8832369It's this dataset https://huggingface.co/datasets/Ejafa/ye-pop which itself is based on this dataset https://laion.ai/blog/laion-pop/ which is 600k ai slop images sourced from midjourney
>>8832374>"600k ai slop images sourced from midjourney">me when I spread misinformation on the internetread the link you posted again, more carefully this time, you illiterate ape
>>8832377sorry, it's only slop images that are similar to slop images on midjourney. huge difference here folks. add ye-pop to your prompt now
>>8832367>>8832370I just use>worst quality, censored, mosaic censoring, bar censorHad no idea ye-pop and deviantart were tags
>>8832378>huge difference here folksIt is unironically a huge difference. It's not synthetic images, it's real internet images that roughly align with the types of things people generate on midjourney.I don't fucking get it. People have been saying shit like "man I wish Noob was better at more traditional and painterly styles in addition to anime". Somebody actually trains a model specifically to do that, and you complain.
>>8832383>People have been saying shit like "man I wish Noob was better at more traditional and painterly styles in addition to anime"People also love shit like WAI, because the people are retarded.
>>8832383>People have been saying shit like "man I wish Noob was better at more traditional and painterly styles in addition to anime"literally fucking who?
i like painterly btw
>>8832385same people that are once again advocating for farting on pray to make its way into anima
>>8832386where is that clown btw, nigga owe me pokesluts
>>8832387every furry I've seen complaining about anima has been trying to get the dev to stop training on deviantart and laion pop, and replace them with e621 instead. it's not the same people, furries only want e621 and nothing else because they are degenerate brainrotted gooners
This nigga is still crying about farting_on_prey in 2026
>haha bro>just forget that we ruined the last 3 models bro>get over it bro>just let us try once more bro
its bird schizo
bro is the prey that got farted on
Uh oh eps schizo meltie!
>>8832397>forget that we ruined the last 3 models browhat models?pony was not ruined, it was a shitslop cartoon porn model made for furries, anime was second class citizen to begin withnoob's issues were not due to e621, eps was fine, it was vpred where they started doing weird shitwhat's the third one even?
>>8832403NTAponyv6 was good for its time, but would have been even better if 25% of its dataset wasn't MLP and another 25% western cartoonsponyv7 was a complete failurenoob, it's impossible to say what effect e621 had on it, but it cost $70k to train and probably would have been a bigger leap over illustrious had it not been contaminated by all the furshitchroma anatomy is completely broken for anything other than 1girl standing because half the dataset is e621. it's seen more paws than human hands
>>8832403>pony was not ruined, it was a shitslop cartoon porn model made for furries, anime was second class citizen to begin withwell imagine if astra wasn't a furry and a brony and wrong in the head>noob's issues were not due to e621, eps was fine, it was vpred where they started doing weird shiteps 1.0 was also shit. even if we pretend that the e621 dataset had no negative influence, its addition still prompted them to unfreeze the te which ruined the bake>what's the third one even?nai v4.5 also has a furry dataset
>muh unfrozen TEOh it's really the eps schizo doing his usual armchair baker routine.
>>8832244>>8832248It's, uh, something https://litter.catbox.moe/x6nc8y4x1wdy3hxt.jpg
unc really boxing his little saint james controlnet slop :skull:
>>8832412>soul>soulless>MAXIMVM SOVL OVERDRIVE
at least e621 gives you good sex and uncensored genitals knowledgewhat the fuck does deviantart and ye-pop give you?
been playing with animait's better than illustrious base but that isn't saying much compared to all the fine tunes. Unironically wai is massively ahead still.It can be potential model just waiting for fix #13123 for now
>>8832417good vibes, you wouldn't get it chud
>>8832418>Unironically wai is massively ahead stillThis, so much this!
>>8832418>Unironically wai is massively ahead still.what's your opinion on dalits, sir?
>>8832319Still waiting a full polished gen using this super duper method
>>8832421>>8832420nooo you can't heckin like the most popular model everything popular is indian!!!
>>8832424Yeah.
>>8832424It kinda is when it comes to the internet and AIslop in particular.
>>8832418In case you didn't know, wai is just noobai eps (aka illustrious) with a bunch of loras merged in
>>8832427sure but it's probably the example of the most stable popular anime model right now.Anima won't pick up steam until it consistently makes better images than it, and it has the potential (natural language works far better in it), but isn't there yet at all
>>8832429and i'll add "why should I care what people like most", that means funding and support like controlnets or lora trainers
>>8832429it's only popular because its creator lied that it's not noob-based which allowed using it for civitai onsite gen. that's it.
>>8832429Anima won't pick up steam until it's available on jeetvitai for online generation and some subhuman finds a a way to merge alldisneyprincesses+incase_lora_pony in it
>>8832431why is noob banished from civitai and how do we know it is based on it
>>8832433>why is noob banished from civitaiLicensing>and how do we know it is based on itIt knows concepts that base illustrious has no business knowing.
I dislike frieren, it insists upon its elf
>>8832433it wasn't available at first due to its licensing. its fine now, but wai got the momentum
How come there are only 30 images with "coconut_bra" on danbooru? wasn't this a boomer ecchi staple?
>>8832240>>8831967do I add both of these in a .txt file, change it to toml and drag it on the anima py thing?
I'm just glad I don't have to use regional prompter and that testicle sucking actually works consistently.
>>8832442nah, you reference the file path of the dataset.toml in the arguments and you can paste the arguments directly into the cmd window while the venv is active.
>>8832443pretty nice, box?
>>8832447>image pulls the viewer's gaze to male pubic hair in very homoerotic fashion>pretty niceuhhhhh, anon?
>>8832445ok so:first I make a toml file from copy pasting this: >>8832240and then change all the directory of this: >>8831967finally copy paste of the latter's text after making sure to activate venv?
>>8832448you're gay
>>8832449yes. remove all backslashes and linebreaks as well if they get copied over into the console. and make sure that the file paths use /, not \
>>8832447https://files.catbox.moe/l30hbe.pngI don't know if it's the model itself or my settings but black backgrounds end up being washed out, though it's easy to fix.
>>8832454thanks
>>8832329After playing around for a bit it seems the default of ER SDE + Beta absolutely destroys styles compared to Euler a + Normal, which replicates them far better.
>>8832468proof?
>>8832468>Euler aAs usual makes everything le heckin smooth. works for some styles i guess
I hate that my beloved DPM++ 2M SDE doesn't work
>>8832469Left is Euler a + Normal, right is ER SDE + Beta, the target artist is Yunsang. You can certainly see his shading on the left even though the style isn't 100% on point whereas the right is just generic slop completely unrecognizable as him.
>>8832294>>8832329512px should be equivalent to 1024 on SDXL, because it's trained at half VAE compression.
>>8832478"should"?
>>8832484Well it's probably not quite the same effect. Seeing as it already produces better eyes and other details on base res gens.
seedance 2.0 hentai finetune when???
>>8832504>seedmaybe as a reward after ten years
>>8832454can you repost the box please?
>>8832468Wouldn't say that's an absolute. ER SDE (I prefer dpm++ 2m over it although they're similar) helps with styles that have more detail or sketchy. Euler A smooths things out a bit which can help with anatomy issues the model currently has but you lose plenty of detail. But as I posted here >>8832282 it absolutely is the best sampler for anything leaning towards 2.5d and overall smoother textures. This anon >>8832331 is right in that currently samplers and pos/neg tags overwhelmingly influence your gen. Which it's nice to have the option but requires extra autism.
How the fuck do I upscale anima models, they all give me broken images
>>8832543read fucker, we already talked about it twice in just this thread
>>8832543Some anons have mentioned boomer methods like ultimate sd upscale. Last time I tried that, ages ago, I kept getting seams all over and a few anons have complained about that with anima so I'm personally just coasting on 1.3x hires until we get some decent cn support.
>>8832545Aw shut up nigger, you read it already so just spit it out like your mom should have done you>>8832546>1.3x hiresSo that's what works? I'm reasonably fine with this actually, considering how it doesn't crush the face and overall details like SDXL used to do
>>8832546you have to do the seam fix pass and stay below 0.4 denoiseI posted a catbox before with working settings, twice
>>8832547>So that's what works?Yeah. Pretty well too. The usual 0.3 - 0.4 range. For where we are now, I can't complain. It looks pretty damn good minus it's still a resolution for ants in 2026.>>8832548I'm willing to give it a try. Where at? catbox seems to be down again.
>>8832549>It looks pretty damn good minus it's still a resolution for ants in 2026I don't care my monitor is 1080p, I only want to upscale so I can add a bit more detail to some areas, particularly eyes since I like to gen them with detailed eyelashes
>>8832549reposting for the third time https://litter.catbox.moe/jy06t8qikipkt6he.pngfrom my earlier tests, hiresfix or any other img2img is fine roughly up to about 2 megapixel (1420 square) before it starts to break down
>>8828594NTA but try something by a famous artist like kishimoto or toriyamaIt's pretty funny
>>8832553Ah yeah. I remember seeing this but didn't put much attention at first as I was still on reforge and just made the move to Neo on Friday. I'll try to parse the spaghetti code later to see what settings you used exactly. Currently still slopping about with tags and samplers.
>Still can't reproduce Nami's towel dropgrim
why does no one ever post anima gens
>>8832561nobody is posting gens in general
Damn this model is atrocious at doing foreskins lmao
Has anyone devised a successful prompt for realistic, accurate anime screencap gens?I've been trying for a while now but it's harder than I thought it'd be
>>8832474https://litter.catbox.moe/pcj7cjgsan5e6gb0.pngIt does. It's just a bit creative as the model page says. Reminds me of genning with AYS on early illu/noob.>>8832563Genitals in general. We need the snakeoil illu penis lora to return.>>8832565>anime screencapunc wildin'
>>8832566artists?
>>8832568It's mainly 'rolua' and 'momfuki rio' but prompt schedule them like this this anon >>8828882 suggested. It's been helping me cope with artist mixing being weak to non-existent right now. And a low dose of @khyle on the side.
>>8832566why is that 1boy so ugly
>1boyWhere?
>>8832570thanks. and i don't think artist mixing being weak is a "right now" kind of thing, we'll probably be stuck with that for the full model as well
>>8832577>returning to lorasGrim. If there was one thing about noob, most styles were achievable (few exceptions sure) with tagging provided your shitmix wasn't slopped to WAI levels. It was nice deleting that giga huge pony lora folder.
>new local model drops >localjeets again stuck training thousands of loras first to make it usablelmao
>>8832584they also complain every day that they have to wait for the full release kekyroooo
bros.. the naisissies are getting courageous again..
>>8832474yes it does
naibabs have been waiting months for v5 and they still do not have vidgens :^(
>>8832282Can you post catboxed example of 'nyl\fugtrup' '3d'? I've tried to gen, but failed.
>bidgens loll
what's the point of having three separate threads if they're all dead?
>>8832646Post an anima gen, it's that simple
>>8832646cause i don't want to look at /pol/ /v/ tard cross spam links, poo poo coprophagia gens, wai base sexy brown style gens
>>8832649i didn't ask how you shitpost that thread bro :skull:
>>8832524https://files.catbox.moe/i41c9i.pngI don't know why it broke
>>8832677thanks, was this inpainted? just asking because i only see robin prompted and not firefly
>>8832647Based anima requester.>>8832617It's far from a 1:1 but considering the limitation of the preview model and that it's not even supposed to do 3d, it's a decent imitation. Use the same sampler/schedule as those do the 2.5 look best.https://files.catbox.moe/008fqo.png
>>8832695Love the subtle hint of incase here, very nice.
>try ada wong>ambiguous looking caucasian cosplayer on tiktok >prompt asian>actually ada>tag is so strong some gens felt like racial caricatures>kekedI kneel, anima.https://files.catbox.moe/48plns.png
>>8832738something is wrong with the character replication on the anima preview
whats with the weirdo behavior of editing your metadata to hide prompt?
>>8832752it's cathag
>you can literally just tell the model what you want>ARGH HELP ME PROMPT NIGGERMAN I'M GOING INSANEYou should really just pay the 25 at this point, unc.
why is scathag like this?
>The composition utilizes the golden ratio to position the figure against the vast urban sunset, creating a powerful silhouette that speaks to ambition and reflection. Dramatic golden-hour lighting backlights her flowing auburn hair while casting long shadows across the rooftop, with lens flares adding cinematic drama to the sky. Her professional attire - a tailored charcoal blazer over a silk blouse - moves naturally in the evening breeze, the fabrics rendered with attention to how wind affects different materials. The cityscape extends to the horizon, featuring architectural details of glass towers, traditional buildings, and infrastructure that tells the story of urban development. The artistic approach combines architectural photography principles with character-focused narrative illustrationThere you go, lil bro. Just replace the words like mad libs.
why is lil blud crashing out?
Just put the prompt in the bag, lil bro.
thanks grok
>>8832789c-chadmixanon? is it you?
>>8832790y
>>8832791cumfy gen but inoffensive metadata.
>>8832794yeah i slop edited an image export node w/ formatted string with stealth metadata
https://civitai.com/models/2377376?modelVersionId=2694423>mogs any sdxl slop
>>8832831omaigoddomore WAI-adjacent slop! YES!
>>8832831why do civitai faggots love this aislop style so much
>>8832833also troglodytes on /hdg/
>>8832831Wouldn't merges with LoRAs using barely captioned images like that destroy Anima's NL comprehension?
>>8832844who cares as long it looks sexy
>>8832844You don't need NL for 1girl, standing
>silhouette sex is a fucking breeze nowbros.. how did we get away without dropping 25?https://litter.catbox.moe/65ishaqy6dfcvcrg.png>>8832844Don't ask questions. Just consume slop then get excited for next slop.
>>8832874Oh that's nice but I will still wait for the full release
>full releasedid nobody tell lil bro?
>>8832872>2girls>try to describe each girl's features separately>horrible bleeding that's only (semi)fixable with regional promptingNo. Never again. Anima is literally perfect for being able to describe the general scene and composition with booru tags and then adding smaller details with natural language.
>>8832876NTA but tell what?
>>8832878When was the last time a trainer improved on the initial release of a model?`
>>8832879Probably Wan 2.2
>>8832879Noob, shit was raw as fuck till full release.
>>8832879Almost every version of noob until the 1.0 were dogshit full of artifacts or not even working at all
model arena from noobai discord>Error: Our system thinks your post is spam. Please reformat and try again.
>>8832689https://files.catbox.moe/l30hbe.pngMy bad, I am retarded
apparently some cloudflare domains are on the 4chan's list??https://hamilton-thumbnails-assess-background.<try>cloudlare.com
>>8832895real shit?
>Do some idle prompting>For some reason the dude looks mildly worriedI wonder if there are any stories out there that start as your usual can't win against cock story just for the dude to wonder if he went too far and regretting the entire thing.
>>8832874Huh. Little effort too. I'm impressed. Still needs some improvements, but this is quite promising.
gacha whore netorare has never been so easy!
>>8832914ignoring the horrid style yeah, was that pure prompt?
anima can't do canariDOAOA
just... train a lora
>>8832915Yes, I only changed Wise's arm to be less muscular, but it could have probably not been necessary with less lazy prompting.
>>8832922I'm not complaining about my random slopped style being called bad, why are you?
>>8832921Impressive, very nice, it's not something that I couldn't do on current SDXL but I am all in for less RNG on my gens>>8832922Stfu n-word
>>8832922>This is why this general is dead.it actually isn't dead enough. A lot of you ahem "people" should go back to /hdg/
>>8832925>it's not something that I couldn't do on current SDXLThat's true for most of the stuff you can do in anima, but as you say the difference is in consistency and not not needing to fiddle with extensions and hacks to force SDXL and CLIP to not be retarded.
I want anima controlnet
>>8832907kino
>>8832932I expect this test bake will not get one, only if the full bake ends up being shit and people come back to this.
how censored is seedance?
>>8833004Literal first result on search engine says they are going to tighten up with the safeguards due to Copyright and other sensitive topics.
>>8832894sorry but the image doesn't even load for me lol
Worth using koikatsu models for a reference image or are loras good enough to make consistent character art?
>>8833029You want koikatsu for pose reference so you can use controlnet over it, you want to use a lora for getting a specific style or character that the model knows very poorly or straight up doesn't know
Man I'm a reforge midwit so I wouldn't know, but autotagger doesn't seem to work on forge neoIt's the only thing stopping me from sticking to it permanently rn
Og SIH but anima shitmix when?
>>8833034What about for OCs?
>>8833049Depends how big if you autism regarding specific body parts, colours and or accessories, you could possibly get it completely or most right from just pure prompt, give it a try
come back :(
>>8833051is your autism*tehe>>8833053to where
So what, I just take my SDXL dataset and just use it for Anima? No strings attached, no changes, no NL shit?
>>8833057yes
any gens?
>>8833067Do you want to see those cute horse girls taking horse dick?
>>8833068that's more of a /trash/ thing no?
Man, idk if it's because I'm used to how sloppy illustrious used to be but all the artists I use seems to have a great (and negative) impact in the anatomy rather than just the renderingWhat artists are you guys using?
>>8833070More or less yeah but since you were willing to see any gen I took the shot
>>8833074yuh, just don't get jannied bro
>>8833045It really is devoid of a lot of features and quite a few extensions don’t work on it that I used on reforge. Nothing major but def a few QoL shit like auto ratio dimension picker come to mind. I’m only using it for anima and reforge for everything else.
>only people left in this thread are animal fuckers, homosexuals with AIDS, and cuckoldsgrimmy!
>>8833082Quite incorrect sir, I am a homosexual WITHOUT aids, please be sure to address me properly
>>8833085you got diddled as a kid and your uncle gave you anal wormssorry lad...
>>8833086I got diddled by porn more likeYou can tell I don't have the worms because I'm a top
>>8833087gays reproduce by molesting kids those repressed memories go hard
>>8833088Having access to unrestricted porn at a young age is similar in spirit to molestationClearly didn't affect me strongly enough though cause I like women too
>>8833089I dont expect you to admit you got anally raped as a childwho would? but im sorry your soul is going to face eternal punishment in hell
>schizos finally killed off hdg so now they're moving herenyo...
>>8833090still believing in semitic fairy tales in the big 26 :skull:
>>8833090That's how I know you're full of shitEvery single faggot I've ever met would straight up just tell me they were molested or raped as a kid
>>8833093im attacking him with the truthno one would just willingly admit they were butt pounded as a kid in this scenario
>>8833094I'm him retard lmaoEvery single faggot I've met, including some I've sodomized, straight up just told me they were touched at some pointThat's partially what led me to be 'bisexual' in porn only
>>8832907Why are they having sex on the windowsill, outside of the bedroom? Shouldn't they be inside and the viewer outside?
>>8833095you still wouldn't admit it here or else you would be mega OWNED and HUMILIATED
>Humiliated>On 4chanlmao
if puttin peanut butter on your willy for the dog is wrong brother i dont want to be right can i get a hell yeah?!
>>8833091chat has been semi decent with anima getting support outside of cumfy too. Shades of pre #10, if I could be so bold. Guess both life support hdg's in the catalog needed to insist upon themselves.
>>8833106the life support must continue so the loonies stay in containment
>nobody:>noodlekek: heres another 1girl standing gen
>>8833106people were actually giving useful lora training advice before #10 so absolutely not
>>8833111Must you smirch him when he has done no wrong?
>>88330194th time's the charmhttps://litter.catbox.moe/aizeusgcc0ws8ksq.png
>>8833115thanks a bunch!
I feel gooning to llm prepared me well for anima's NL support. Like I know how to write to this little qwen lobotomized nigga, chat.
>>8833117not on anima sadly
>>8833071Base illu isn't sloppy. Only the later versions, and merges on top of illu. I like the anatomical impact. But score tags negate it to some degree, and you should be able to avoid it using prompt scheduling. Overall body shape is determined in the very early steps.
>>8833111let me guess, you need more?
poop goes to the other thread
I don't want to know the exact consistency of your feces just go away
>>8833123>I like the anatomical impactI do too but some styles I had worked great before and now make it all wonky. I suppose prompt editing is the way to go.>But score tags negate it to some degreeAny consensus on the score tag business btw? Not that this place has ever reached any consensus before but it doesn't hurt to ask
>>8833148I'm just using masterpiece; worst qualitynot like you need any at all when using an artist prompt, just out of habit
Any guesses what loras pic related is using to gen images?
>>8833158it's nai
>>8833158mmm that sweet one single background WAI is capable of...
>>8833158idk, post more huge milkers like this
>>8833158it's slop
>>8833158Looks like Taimanin Asagi CG lora.
chat, anima full eta?
>>88331792 more
>>8833183i can't...
>>8833192just play any game
do illustrious trained loras not work on noob? I was trying out https://civitai.com/models/2363696
>>8833195any game sucks
Anyone here using Kohya GUI to make Anima loras? I cannot for the life of me figure out how I am supposed to configure it. Could someone please share a configuration file on how to use it?
Just a heads up to my fellow reforge refugee bros. All upscalers now go in the ESGRAN folder, no matter the class. Was malding thinking it was yet another thing Neo didn't support but fortunately did a bit of github reading and saw they merged all upscalers into 1 folder.https://files.catbox.moe/othesi.png
you have AIDS
>>8827652FRIENDLY REMINDER THAT IF YOU USE ANY OF THESE OVERCOOKED PIECES OF SHIT, THE BEST THING YOU CAN DO IS EUTHANIZE YOURSELF AND DONATE YOUR ORGANS AND GPU TO PEOPLE WHO NEED THEM.
>illu and noob models in Mashallah's 2026What did lil bro mean, chat?
It's the year of the Horse, where are the pics of Yixuan shifu getting gaped by horse cock?
>>8833327this is a message for /hdg/
Too busy repopulating the tiger thiren population
>>8833327based and real
>chinkslop
what lil bro yappin' bout this time
>>8833199Works, slightly weaker. By how much depends on the lora case by case. But I'm not sure rectal flow is still considered noob in this regard, perhaps he trained it too much in a different direction.
>>8833335yellow post
>>8833128box please?
>>8833335umas getting horse cock is the real deal
Has anybody trained or found a decent anima style lora yet? It's really hard to get consistent styles with just artist prompts.
>>8833499I'll train you one if you have a dataset
>>8833503I'm already training them myself, just having trouble getting something I'm super satisfied with. If you have parameter recommendations that would also help.
>>8833504that barebones kohya config gave me a pretty good result for a first try.
>>8833506Can you post the lora?
>>8833507yeah in 4 hours or so
Wish my PC was good enough to train loras
i've been proompting too hard, i need to train...
>>8831967i'm having trouble copy pasting this, I keep having this as an example:At line:1 char:3+ --vae_chunk_size=64+ ~Missing expression after unary operator '--'.At line:1 char:3+ --vae_chunk_size=64+ ~~~~~~~~~~~~~~~~~Unexpected token 'vae_chunk_size=64' in expression or statement. + CategoryInfo : ParserError: (:) [], ParentContainsErrorRecordException + FullyQualifiedErrorId : MissingExpressionAfterOperatorI tried removing line breaks, removing \, changing \ to ^, making everything in a single line and nothing worked. I also have venv activated with the green (venv) on sd scripts.I also made a toml already from the same guy posting it. and searching posts about it and nothing came up.
>>8833571>green (venv)what? win11 got colors in cmd or something?try pasting this with updated file paths: accelerate launch --num_cpu_threads_per_process 1 anima_train_network.py --pretrained_model_name_or_path="E:/sd-webui-forge-neo21/models/Stable-diffusion/anima-preview.safetensors" --qwen3="E:/sd-webui-forge-neo21/models/text_encoder/qwen_3_06b_base.safetensors" --vae="E:/sd-webui-forge-neo21/models/VAE/qwen_image_vae.safetensors" --dataset_config="E:/sd-scripts/config/dataset.toml" --output_dir="E:/sd-scripts/output" --output_name="anima_test" --save_model_as=safetensors --network_module=networks.lora_anima --network_dim=16 --learning_rate=1e-4 --optimizer_type="AdamW" --lr_scheduler="cosine" --timestep_sampling="sigmoid" --discrete_flow_shift=2.0 --max_train_epochs=10 --save_every_n_epochs=1 --mixed_precision="bf16" --gradient_checkpointing --cache_latents --vae_chunk_size=64 --vae_disable_cache
>>8833571on that note, I also have all the paths added in too, with the selected safetensors starting from the drive via copy as path.
>>8833571sounds like maybe your qwen unet has a weird file path that breaks the parsingif you think it's the backslashes you can just remove and put everything on one lineoh yeah nta
>>8833580>>8833577almost there, now it's giving this: Error on parsing TOML config file. Please check the format. / TOML config_util.py:693 形式の設定ファイルの読み込みに失敗しました。文法が正しいか確認してくださ い。: D:\kohya\sdtoml\Anima.tomlTraceback (most recent call last): File "D:\kohya\sd-scripts\venv\Lib\site-packages\toml\decoder.py", line 511, in loads ret = decoder.load_line(line, currentlevel, multikey, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\kohya\sd-scripts\venv\Lib\site-packages\toml\decoder.py", line 778, in load_line value, vtype = self.load_value(pair[1], strictly_valid) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\kohya\sd-scripts\venv\Lib\site-packages\toml\decoder.py", line 866, in load_value raise ValueError("Reserved escape sequence used")ValueError: Reserved escape sequence usedthanks guys.
reminder that if you're too retarded to use cmd then easyscripts just workshttps://github.com/67372a/LoRA_Easy_Training_Scripts/tree/refresh
>>8833593>Please check the formatWell, did you check the format? it's clearly moonrunes instead of a normal peoples language.
>>8833595i'll try this one out then, thanks.
>>8833593The formatting is wrong, should look like thishttps://litter.catbox.moe/yzu38xrr7klley0u.toml
>>8833597this isn't the exact error i got but there was an error in one version of sd-scripts caused by a bad thing that's supposed to automatically swap between jp or eng depending on system lang, it hits the jp and breaks. fixed it by deleting the jp text in the affected python files
>>8833602that could exactly be it, my pc is set with jp local and unicode to avoid mojibakes. so i'll try out the ez scripts and see if it works.
haven't been here for a while, what happened to xenoblade anon?did he die before his wish for a 16ch vae was finally answered?
>>8833615i'm sure he'll be backthe siren song of exact character designs calls
>>8833595 got it installed, set everything up now it's giving an error for trying to find a backend.
>>8833627post the error
>>8833627blud you're kinda cooked if you wanna use python but can't even fix simple errors
just ask chatgpt about the error like everyone else :^)
i cant figure out how to hiresfix with anima, im actually using noob 1.1 w/ it as an img2img model for upscale denoise and inpaintinganima does ok with inpainting but seems to need a lot more denoise%
>>8833635doesn't exist rn unfortunately
blud is upscaling a 16ch model with a 4ch model
>>8833627did you answer "y" when it asked if you're using it locally?
>>8833635use tiled upscaling with ultimate SDor only hiresfix up to about 1.3x
>>8833649no i mean, with only a few steps of denoise on an img2img upscale pass all the sampler/scheduler combos i tried just added artifacts and shittiness
>>8833653Yes, because you're going too high resolution. Do not go above 1.3x base res (2 megapixel) or use tiled upscaling.
well i finally figured out what kind of retardation was keeping me from being able to install that ezscripts fork so time to start testing training it seems
>>8833654actually good tiled upscaling needs CN model which we do not have
schizo negs can actually work in anima. try this at the end of your negatives if your style mix is trending too much into westshit:a western style image with pronounced noses and lips.
>>8833656Yes well, better than nothing. 0.3 or less denoise and with seam fix pass is the best we have rn, just takes a long time.
>>8833657people who don't put western trash in their prompt don't have this problem
>>8833660i use only nippon artists, the ones with detailed skin and light can also trend into ugly noses. it helps in that case
Some nigga on huggingface is saying that replacing the default TE with this https://huggingface.co/DavidAU/Qwen3-0.6B-heretic-abliterated-uncensored/tree/main apparently makes it adhere better to nsfw NL.
>>8833627it's giving this error:Traceback (most recent call last): File "C:\Users\ThisPC\AppData\Local\Programs\Python\Python312\Lib\threading.py", line 1052, in _bootstrap_inner self.run() File "C:\Users\ThisPC\AppData\Local\Programs\Python\Python312\Lib\threading.py", line 989, in run self._target(*self._args, **self._kwargs) File "D:\LoRA_Easy_Training_Scripts\main_ui_files\MainUI.py", line 189, in start_training_thread self.train_helper(url, Path("queue_store/temp.toml")) File "D:\LoRA_Easy_Training_Scripts\main_ui_files\MainUI.py", line 197, in train_helper response = requests.post(f"{url}/validate", json=True, data=json.dumps(final_args)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\LoRA_Easy_Training_Scripts\venv\Lib\site-packages\requests\api.py", line 115, in post return request("post", url, data=data, json=json, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\LoRA_Easy_Training_Scripts\venv\Lib\site-packages\requests\api.py", line 59, in request return session.request(method=method, url=url, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\LoRA_Easy_Training_Scripts\venv\Lib\site-packages\requests\sessions.py", line 575, in request prep = self.prepare_request(req) ^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\LoRA_Easy_Training_Scripts\venv\Lib\site-packages\requests\sessions.py", line 484, in prepare_request p.prepare( File "D:\LoRA_Easy_Training_Scripts\venv\Lib\site-packages\requests\models.py", line 367, in prepare self.prepare_url(url, params) File "D:\LoRA_Easy_Training_Scripts\venv\Lib\site-packages\requests\models.py", line 438, in prepare_url raise MissingSchema(requests.exceptions.MissingSchema: Invalid URL '/validate': No scheme supplied. Perhaps you meant https:///validate?I don't get it, comfy, kohya gui and pancho sd installed and works fine but never these.
>>8833670that's bullshit but i believe it
>>8833670Which one am I supposed to download there's too many
>>8833676the heretical one, that makes it extra nsfw
>>8833685proof?
>>8833687dude
>>8833687just making a joke
kekmao
>>8833674Have you tried following the installation instructions?run the install.bat or add "run_local": true to config.json
>>8833696thats set true in the config. with the backend url set to blank.and read the install plenty of times and it'll still end up with the same error. the gui seems to work it's just until I press start training. after that stop training changes from the start and it'll not work and pressing stop gives the same error.
>>8833706>with the backend url set to blankset it to 127.0.0.1:8000
>>8833707it gave this now: File "D:\LoRA_Easy_Training_Scripts\main_ui_files\MainUI.py", line 197, in train_helper response = requests.post(f"{url}/validate", json=True, data=json.dumps(final_args)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\LoRA_Easy_Training_Scripts\venv\Lib\site-packages\requests\api.py", line 115, in post return request("post", url, data=data, json=json, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\LoRA_Easy_Training_Scripts\venv\Lib\site-packages\requests\api.py", line 59, in request return session.request(method=method, url=url, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\LoRA_Easy_Training_Scripts\venv\Lib\site-packages\requests\sessions.py", line 589, in request resp = self.send(prep, **send_kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\LoRA_Easy_Training_Scripts\venv\Lib\site-packages\requests\sessions.py", line 697, in send adapter = self.get_adapter(url=request.url) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\LoRA_Easy_Training_Scripts\venv\Lib\site-packages\requests\sessions.py", line 792, in get_adapter raise InvalidSchema(f"No connection adapters were found for {url!r}")requests.exceptions.InvalidSchema: No connection adapters were found for '127.0.0.1:8000/validate'
>>8833711my bad, it's http://127.0.0.1:8000
>>8833712fuck it's finally running for once, thanksoddly enough it's only using 8 vram
>>8833715it's a small model
>>8833716That's what I told her but she didn't understand :(
>>8833670Anybody try this yet? I want somebody else to waste their time before me.
>>8833725I'm not really noticing any difference t b h
>>8833670oh, nsfw NL? doesn't seem to affect tags much. and it craps on gen quality and character knowledge
>>8833751it's why you don't train the TE in a finetunethough this is a step worse, finetune with one TE then replace it so they're not aligned anymore
are any of the chink or jap anima shitmixes on civitai usable?
>>8833772the full model isn't even out yet so everything that is being mixed or working on it's not going to make a significant difference against the current state of the model
>>8833774i mean illu0.1 had some usable loras and shitmixes despite it being pretty raw
>>8833775but illu was built over SDXL which was something people already had a hefty experience working with
>>8833430https://files.catbox.moe/74wvda.png
Interesting note on the Anima upscaling issue https://github.com/Comfy-Org/ComfyUI/issues/12438
any snake oils you found for anima to not fry with cfg?
made an anima artist style listinghttps://gitgud.io/gayshit/anima-artists-promptsnot really a replacement for the other once since i have less artists but it existsi plan on adding a third prompt more aimed to nsfw and maybe some stuff similar to tagexplorer
>lora key not loaded: layers 0-27uh oh, am I fucking up somewhere?Also is it the same 3000 steps for styles?
looked at what i baked overnight and training on anima is great, really picked up the finer details better than illu. gotta do more testing but it's really nice. i'm also seeing fairly minimal concept lora style influence, or maybe it's just that artists work really well in anima
>>8833975So, the standard kohya config just werks?
>>8833983i ported over all of my settings from before and it works well
Any guides where I can learn the basics for comfyUI to gen animæ?I basically only know how to install custom nodes and change values. a1111 user btw
geg?
>>8834029There's too many guides everywhere, but nothing thread-approved. (And there's not much of a thread left to approve anything)I can answer specific questions.
>>8834042I see.Well my first thought when using comfy was that I need a node to see metadata, like PNG info or whatever is called in Forge. Second step would be, once installed, how to insert/add that node into the workflow.There has to be somewhere out there a compilation of essential or basic nodes with features that the forge UI has or similar. I feel like an utter brainlet again.
>>8834044anima works on forge.
>>8834045Not on my pc and therefore I'm forced to use comfy.
How much of a difference does flash attention do in lora baking? Anima takes so long at an hour above, where I can bake noob loras in less than 30 mins.And holy shit the trouble you have to do to attempt to even install it.
>>8834071i haven't tried it. anima increased my typical ~1:50 bake time to ~2:30, it's not a huge deal
>>8834071have you tried using glora/unet only in anima? that's the method I usually use and wonder if it bakes the same. my first baking in anima never really worked so i'm trying to copy my best via metadata.
>>8834044Not sure if there is a good one, I use https://sprites.neocities.org/metadata/viewerMany images you can also just drag into the UI. For comfy this imports the workflow, for simple A1111 stuff it will try to produce one (though it will set incorrect clip skip on SDXL, because it's saved incorrectly in webUI)Here's mine with all the common features, labeled and annotated. Using the bare minimum custom nodes.https://files.catbox.moe/w67rvs.png
>>8833834This is great
>>8834120ngl almost every artist i've found works awesome in anima
WTF when did video models get this good>https://x.com/senyokahi/status/2024119233135673635?s=20
>>8834162like...3 months ago? you may have missed it but the first big round of models that are trained with ai helping to write the code that trains the model are coming out. shit is accelerating
>Ask claude to put @ infront of the artists in the auto complete csv>this nigga builds an entire app including drag and drop file uploadpretty sick ngl
>>8834176Cool, I wonder if I can make a search tab to find artistsSomething that allows you to filter by tags, date and post count would be awesome
>>8834122I think it's a bit inconsistent though
>>8834230i mean it's undertrained and 512x
>>8834233What does it even mean that it's 512x?Btw I'm not sure the inconsistency is a defect per se. When it comes to style mixing, for sure, but when it comes to a single artist then I think the inconsistency is a byproduct of how good the prompt adherence is.Y'know how in lora training if you feed the model several arts of the same artist, but with different artstyles it'll produce a bad lora, so the goal is to give it only the same consistent artstyle? With how well the model sticks to a tag, I think it acts kinda like that. So artists that have several distinct artstyles will get inconsistent results, or if they're constantly tied to the same franchise (like with big artists; kubo, oda, kishimoto, toriyama), or if they have enough of posts to have his development as an artist recorded.
>>8833834oh https://thetacursed.github.io/Anima-Style-Explorer/index.html changed the prompt but now it only has a blank white background so you cant see how well the artists do backgroundsalso he has a fucking donation link at the top.>>8834243trained at 512x512 instead of 1024x1024 so it's missing small details during training though it's still better off then SDXL because of the improved vae, also why it has trouble upscalinhdont quote me on this though i dont finetune shit
>>8834247I don't think it's the reason upscaling is fucked, SD1.5 was trained in 512 and has no issues upscaling.
>>8834250I don't remember hrfix being a thing until after NAI leak which was trained on slightly higher res but it's been 30 years since then so I could be misremembering.
>>8834247Wait, so basically it's trained on shit resolution images but since the VAE is better it still looks vastly superior to SDXL without adetailer/upscale?Insane
>>8834251I mean it can hiresfix up to 8K, as long as you keep the denoise low so it doesn't have enough leeway to fuck up anatomy. I actually downloaded based64 again just to check. Anima starts breaking apart at 2mpix and makes square artifacts, regardless of how low your denoise is.
>>8834253It only does half as much VAE compression.
>>8834253if im understanding the anima page correctly, yeah, and they're going to train it at higher res later>>8834256good point, sd1.5 never quite broke in the same way anima does while upscaling
>>8834253that's why it's so hype
Chat, your name is not in the files, right...?You wouldn't lie to (me), right?
>>8834283fantasies should stay fantasies
Feels like an otoko_no_ko (now trap because anima has updated danbooru tags) kinda night.
Can someone else use @moreshan and see if it works on Anima? I don't think it reproduces his artstyle correctly at all even though he has 270+ posts
>>8834247>he has a fucking donation link at the topwhat a grifting fag
>>8834341that's not what grifting means...
>>8833834Mostly finished vibecoding the site for this. https://anima-artists-prompts-957be3.gitgud.site/You can search for artists and if you click on an image it copies the artist name to your clipboard.It doesn't have a donation link so it's at least 0.3% better than the other dude's.>>8834324It kind of resembles his style if you add the "one piece, one piece card game" tags but that bakes in the card game logo obviously.
I miss artist mixing>>8833834>>8834367>i plan on adding a third prompt more aimed to nsfwwould be greatI think there is a lot more to a style than just the "style", an artist style affects so much more like how poses are set up, composition, anatomy and body typesI think what the other guy has done with just having a girl facing the viewer with a white background is a mistake as it purely showcases the base style and nothing else, it's not very useful unless you're only interested in knowing how an artist draws their faces
>>8834396yeah, if there's no action or pose or backgrounds or opportunity to show off lighting kinda lame
>>8834396yeah, I didn't like either of his choices for representative prompts. i tried to make a prompt that was a bit more general with the rosemi prompt, showing full body proportions, facial proportions, and backgrounds all at once while the minimal prompt shows what the artist is most naturally biased to outputi was thinking of something like this for the nsfw images but this pose doesnt show anus so i'll have to workshop italso you can mix artists its just really wonky, feels like bringing weight up is more effective when mixing then bringing weight down
>>8834400that or extremely low weights like 0.2
>>8834396>I miss artist mixingJust put the scheduling in the prompt, lil bro.
>>8834406scheduling works rather well but it's not quite the same as mixing
>>8834406put a bullet in your brain retarded zoomer
unc crashing out bigly
>>8834406Put the scheduling in the prompt and no one gets hurt!
>>8834406What's the ratio? 0.5?
>>8834367You should add sorting by post count or filtering by post count imoBut it's pretty good>It kind of resembles his style if you add the "one piece, one piece card game" tags but that bakes in the card game logo obviously.Yeah, I think that artists that are heavily associated with copyright tags don't work properly>>8834396I miss it too, haven't found a single artist that fulfills all I want
>5 threads before we fall of the edge of the worldnoodlefriend save us
>>8834446On it
Alright, move>>8834450>>8834450>>8834450
>>8834451>/aco/ op
