>>8893003 yeah 300 gens across three months in two different threads because you had to bake a new one since nobody was posting in it is quite impressive
>>8893018 then you should tell the retard(s) keeping that place on life support that instead of telling people posting in this thread to go back. actually, don't do that.
imagine your thread being so dead you hit page 10 at 150 images and are forced to make a new one, then proceed to beg the people in the other general to come back
desu, i prefer the name /hgg/ over /hdg/. /hgg/ sounds like you're trying to push a fat stinky dump that isn't comming. like hgggggggg!!!!!! quite fitting considering what /hgg/ favorite food is
this place would have been so much better if it required at least 200 characters for a text-only post (either 0 or 200+ characters with image attached)
>>8893056 there's tanaka makoto and panzermeido - not born as israeli but both have lived there most of their lives. i can't find any palestinian artists on danbooru.
>>8893066 How does it feel knowing that the useless weak seed you spawn will just be wasted on a pillow or on the hairy belly of some ugly fat homo bastard?
>>8893135 >>score_7 It's part of the officially recommended Anima pos/neg sloppa. >also have you considered maybe you're retarded and that "nemlei \(style\)" should be "@nemlei \(style\)" ? No, you. nemlei \(style\) is a generic blue tag on the booru, not a red artist one.
>>8893130 this is what local gets for completely misunderstanding and disregarding cfg and how to use it as well as other commonly applied training techniques not even touching dataset processing
>>8893162 i don't get why people skip research and go straight to training like, actually understanding what went into the models that came before your own
>https://x.com/NatsuMegu_7210/status/2062142630939828252 >THIS flavor of ai slop is still blowing up on twitter >(You)r posts are flopping why might this be?
>>8893210 never was, restarts serve no point if you aren't saving the intermediate checkpoints, and even then at best they don't achieve anything while slowing down training, at worse they also hurt the quality
>>8893246 interesting someone took the mantle of asking how to bake including screenshots. I really hope someone advices something new that's better than came/rawr combo. that's so much pics to bake hope it goes well. This is bat 4, but only 32-45 images at typically 90-125 epochs.
>>8893269 I have been using his config and in dataset his shuffles and doesnt have caption or tag dropout, also has random crop padding at 0.05. Its bad?
>>8893342 >we can ignore you little brown bro is that what you're doing right now, queer little princess? so far you're acting like a girl with a secret crush on me >j-just so you know, i'm gonna ignore you!! that's right, i'm ignoring you, MISTER!... what do you have to say about that?.......... HEY!! act your age, fatty. your bring shame to your mother.
>>8893371 >fingernails, toenails, (nail polish, toenail polish,:-1) These work for me. If you just use nails they will always come out colored, if you put nail polish in negpip they will usually come out colorless.
found a node repo where someone got some tile segs thing to upscale with anima claiming it's better than seedvr2, but i can't seem to get it working it just produces a blurry mess. is there any legitimacy to this approach or should i just stop and work with something else?
>>8893379 try adding in the anytest controlnet as well and see how it goes. my preferred method of upscale back on sdxl was tile with cn and regions i manually set up to avoid hallucinations. haven't tried it on anima tho
>>8893381 Thats not the problem. The problem is there is no tag for "unpainted" nails. If you specify sharp nails, it has tendency to add color. But there is little way to get rid of the color.
>>8893442 Retard this is literally the example picture https://danbooru.donmai.us/posts/5030937 >Having fingrenails does not imply they are unpainted. Unless you tell the model to color them they won't be unless it's some character that already comes with painted nails by default
>>8893444 https://danbooru.donmai.us/posts?page=999&tags=fingernails+nail_polish+ There is way over 1000+ pages that combine fingernails and nail_polish. So having just fingernails does not imply no nail polish.
>>8893130 two factors in this. One is that anima isn't super familiar with the nemlei tag due to it not being common on danbooru. The other is that the tag "official style" is making it copy the azur lane artstyle for the azur lane character you're trying to gen. The solutions to this are to either go on https://tagexplorer.github.io/#/artists to find a similar style, make a lora if you really want that particular style, or just stick to nai if you're fine paying money
>>8893565 NAI and Anima were trained on largely the same dataset, it can't possibly be the fault point. Any differences in booru prompt comprehension MUST be caused by something else.
>>8893568 nta, I think the same applies to other franchises as well. In general tags leak less style than they used to on sdxl, if they don't explicitly start with @.
Like if I do "1girl, blue eyes, blonde hair, Senran Kagura", noob gives me a generic girl in the game's 3D style, while anima gives me Yomi in some generic AI style.
>>8893612 it's crazy how much more appealing horse anus is. both make my mouth water, but for black person anus, not in a good way. same feeling as how your mouth waters when you're about to puke
idiot fool impostaar. everyone knows black people identify themselves using their anus, it's their version of fingerprints technology, inspired by dogs and cats
>>8893692 he always spazzes out when there are too many text posts and no one is posting anything. I think the real annoying part is our dumbass janny ignoring blatant acoshit.
All of my hard working shilling NAI as the next best thing after being White... years I will never get back... all of it down the drain, ruined by a single person with $25 and a compulsion to keep clicking that generate button. NOOOOOOOOOOOOOO
>>8893715 Looks like a sheathed horse penis anon. In some centaur lores the horse part is a monster that fuses with a human female. Basically eats her legs and melds their bodies together via the butt/birth canal. I imagine this process would explain how this girl has both anatomies.
>>8893725 Horses and many other mammals have a sachet of skin that the penis recesses into when flaccid. Soft vascular flesh like that would be an easy priority target for opportunistic predators just waiting for Chad mcstud to walk around with his 30 inch penis just hanging out all the time, and you can't reproduce easily with a damaged dong.
>>8893733 >>8893736 >>8893738 Back when you could spoof infinite free nai trials I learned how to abuse the generator pretty well. Blame comfy for hating my GPU and Nvidia for ruining the pc market.
>>8893267 the 1million dim lokr is about the same as rawr/came standard, but it has a harder time doing darker darks and likes more of a soft light (orangish/yellowish) when trying dim lights/dark, I still think normal locon is better, at least the 32/32 combo, still it was a good try I expected something deepfry.
>>8894001 I need to clean out the majority of mine although there's still a few loras I have to hang on to (alongside a noob model) as anima lora support is rather slow in the wild and some of the illu/noob loras I don't see foresee anyone baking on anima.
>>8894025 I'm surprised /s/ doesn't have a diffusion general. If you create one, let me know. I'll post some stuff there to prop it up. I just can't be arsed to bake the first thread.
>>8894055 Why would this kill /s/? It's one thread that barely had any traffic anyway, and it was killed because /aco/ is for adult cartoons and western comics, not realistic ai images. /s/ fits, imo.
>>8894057 /asdg/ was killed because it didn't fit the board it was in (/aco/ is not for realistic images, lol). I don't know about /r/, but I guess it's because that wasn't requests.
>>8894058 /asdg/ existed for years yet was killed a week after they killed /r/. It was for deepfakes, AI porn of real people. Once you're doing photorealism, jannies have no way of knowing if it's someone's face or an OC. So they pussied out on the whole thing.
>>8894065 I haven't seen it in a while, but it was mostly Taylor Swift in terrible images. When it comes to /b/, though, the risk is always there, I guess.
>try an Anima LoRA for the first time now that they're actually being made >bizarre artifacts when Hires. Fixing but not in normal res or if I up the base res Anyone else had this happen? It happens EVERY time, but only with a LoRA. https://files.catbox.moe/9ucqd4.png https://files.catbox.moe/3qlxc4.png
>>8894149 >2xPublic_realplksr_dysample_layernorm_real I had the same issue with this upscaler, exact same kind of random artifacts, I switched back to just using lanczos since it happened even without any lora sometimes.
>>8894172 That's pretty weird because I've been using it for a month and this, with a LoRA, is the very first time I run into it, and very consistently at that.
>>8894173 >try running the upscaler on its own I did and it was the culprit, there were some images that would get fucked up no matter what. >>8894174 It was rare, but it still annoyed me enough to stop using it as the default upscaler.
I can't seem to make a girl stick a dildo up her ass in anima. It always looks like a cock is coming out of her asshole. If she's sitting on it then it usually works ok. What might I be doing wrong?
>>8894178 Damn, going back to 2x-AnimeShrapv4_RCAN really does fix it. Goddamit, I liked the new upscaler better. Oh well. Also apparently CFG++ fucking sucks ass with LoRAs when upscaling so back to good ol' basic Euler a it is.
>loratrannies: "lolmao just train a lora ecks dee" ok >model that was functioning perfectly fine on its own suddenly starts having all sorts of issues from upscaler artifacts to fucked up colors in hiresfix
Wait, we are still doing the soý to onions thing in 2026 and that it's a site-wide thing in the first place instead of being limited to /pol/ and /pol/-infected boards
>>8894380 iirc it actually started on /fit/ used to be just an argument between good plant protein and xenoestrogens, but it never led anywhere productive the schizophreniac theories were tacked on later with the filter already in place
What do people use to NL caption images that would obviously get censorslopped/rejected by mainstream models? TDRussel recommends Gemma 4 31B but it will obviously refuse to caption any images with nudity, let alone sex.
>>8894403 her body is like an ocean world with kilometer high waves for my sorry (shrunken) brown self, and i don't know how to swim. bbllrlrrlrrrrrrpppghhhggppp
ever since the spam of unedited nai images I can't take kurumuz serious anymore. hope v5 isn't such a failure and local can take a break from single-handedly pushing the tech ahead.
i'm still stuck on an illustrious/noob self made shit mix. i feel like i'm the next generation of one of those english cigarettes that was stuck clinging on to the dead bloated carcass of pony...
>>8894541 >>8894542 The style is just >3d, 3d \(artwork\), 3d animation, blender \(artwork\), source filmmaker \(artwork\), red theme, high contrast which gives completely random results, going from that to this https://litter.catbox.moe/nxpuuzli6o48uml7.jpg also, individual 3d artists don't work on anima. So you better train a lora if you want to see a specific style.
ok so, I asked the slop machine to do this but I didn't like it, it is a little too clean, it doesn't feel like the kind of rough animalistic sex it's supposed to be >https://files.catbox.moe/h0ct1g.png
so I thought maybe the perspective and pose was the problem so I change it to this one which is the one I liked the most >https://files.catbox.moe/gm0h9q.png
>>8893883 this is just gay sex. all the girl is doing is letting them use her anus as a container to keep their cawks squished together. 0 sexual stimulation from the female
>>8894666 >give me the metadata, i'd like to try try what lil bud, ruin eveything?, is the same style as the honse gens, characters are zuikaku and cleveland get to work (or not)
>>8894683 well it's post-rough it's dumb anyway cause wolves don't scratch each other for mating, would just reduce his chance of having offspring if her wounds get infected
If anyone's still messing around with vidgen, I've been trying "10Eros", a nsfw finetune of LTX2.3. It's still heavily biased towards 3D and much more so than Wan, but seems to at least understand anime anatomy better now and it's not censor cucked. I'll spend some more time on it, see if it can be convinced into 2D.
Main selling points should be that it's a newer arch and a much bigger model so can hopefully handle more complex interactions. Also you can generate up to 20secs in one go, and optionally at higher framerates and with audio. That said higher framerates are probably not what you want for anime https://litter.catbox.moe/ylg5es3czczz71zc.webm
>>8894696 Kind of a shame that LTX has so little of an idea what anime is supposed to look like even with shitmixing, WAN has way worse motion but at least it attempts to remain in the realm of 2D, if failing to be perfectly consistent style-wise.
>want to keep "extra arms, extra legs, extra eyes" in negslop for obvious benefit >but also want to get rid of them in case I want to gen monster girls that have such features one day What do?
>>8894720 I use WebUI's styles feature so I have to go out of my way to edit pos/neg sloppa and would prefer to keep it as set-and-forget and universal as possible.
>>8894723 >I use WebUI's styles feature so I have to go out of my way to edit pos/neg sloppa what are you talking about, you still have to paste the style in the prompt box with the button, which you can then edit
>>8894727 >you still have to paste the style in the prompt box with the button Not if I set that style as the default in ui-config.json. >which you can then edit Yeah, by clicking a button that opens a new full-screen menu, aka going out of my way.
>>8894740 I thought that just preselected it for copying but didn't actually apply it, good to know. In that case, just press the copy button and edit it in the prompt window for those single gens.
>>8894733 Sure https://litter.catbox.moe/rptvp65bat0q31og.webm But it's just the bare minimum to run LTX. It's not generating audio for some reason so I disabled that branch. I think that might be an issue with the finetune but don't care enough to investigate.
>>8894735 Someone more technical can give a proper explanation. But in simple terms, the longer your negs are the less creative the model will be, in terms of compositions, environments, camera angles, character features that you leave up to chance, seed variability, etc. Because the side-effects and biases from all your negatives add up, even those less obvious or hidden ones. Until the model is guided away from too many things at once.
>if all the tags are actual booru tags They are still being interpreted by an LLM, which may assign them additional meaning you are not aware of. Less likely to than CLIP would, but still.
>>8894750 So basically, instead of only asking yourself "what do I not want to see in the image", you should also ask "what could indirectly be associated with the tags I'm putting in the negatives"? For instance, in >score_1, score_2, score_3, (ye-pop, deviantart:1.2), blurry, worst quality, low quality, cropped, jpeg artifacts, scan artifacts, sepia, monochrome, film grain, chromatic aberration, signature, watermark, username, artist name, web address, patreon, logo, artist logo, artistic error, bad anatomy, bad hands, bad feet, bad face, mutation, deformed, chibi, character doll, extra eyes, extra arms, extra legs, extra digit, fewer digits, missing limb, wrong foot, wrong hand, long neck, cross-eyed, bad proportions, zoom layer, blank page, ai-generated, old, early , the entire block about watermarks should probably be removed because a lot of good artists use these and the model could erroneously associate good images with those tags, meaning you would be reducing quality by putting them in the negs? Because I really don't see how any of the other tags could ever be associated with good images.
>>8894771 Keep your negs empty except for what you need for style. Then when something unwanted shows up in a specific prompt, you add the neg, then remove it afterwards when moving to a different prompt.
>I really don't see how any of the other tags could ever be associated with good images It's not about good and bad images, it's that there are many different kinds of bad images, and you don't know what resulting vector they average into. Yes they all have bad hands, but 80% of them might have more detailed backgrounds than the danbooru average. Like robbutts.
>bake anima character lora >outfits barely work >in a fit of rage, raise lora weights to 3.0 >they work so do i increase the LR or up the step count? this was already 24 epochs (4k steps total batch size 2). i also noticed a lot of anima bakers are just doing separate loras for each outfit, is anima just bad at outfits compared to illust?
>>8894793 lr is 6e-05 >>8894791 maybe later. the character i'm trying to train is kosaka shizuru from taimanin. i might just restart with more balanced images and try to do some crops
>"party horn" didn't really work on SDXL and would only give a party hat and would sometimes give a random straw type thing the character would hold >anima is actually worse and it will ONLY give the hat >and it's doing it largely off-center, as if it were some sort of asymmetric horn fucking lmao
>>8894877 oh but appending "party horn" to "party horn in mouth" with "party hat" in negatives almost works. like, it still doesn't actually work. but it almost tries to do it.
>>8894817 With Wan2.2 it's viable, the worst part is having to switch between two checkpoints when only one at a time fits into vram. With LTX, well this is the halfsize version and with a distill lora. Means you're getting a shitty braindead version of the actual thing.
>>8894775 >Keep your negs empty except for what you need for style. Then when something unwanted shows up in a specific prompt, you add the neg, then remove it afterwards when moving to a different prompt. That sounds a little extreme if I'm being honest. The model already outputs abysmal dogshit with the officially recommended prompt, which is really quite minimal, I can't imagine it doing any better with an empty neg.
I know this will sound dumb anons, but how do you guys actually upscale anima? I just can't get it to look as sharp as my sdxl gens. The closest I've got is just simply doing latent upscale, but anima breaks at 2x and a bit below that. Tried multidiffusion but not good enough. Even in img2img/inpainting, I can't get it to look sharp or inpaint nicely.
>>8894887 It's all theoretical anyway, if you don't mind the outputs you're getting with a schizo neg then of course keep it. I am biased, from spending too much time with noob and using empty negs due to that technical quirk. Which showed me that they're never truly necessary, the model's already avoiding the bad stuff with most positive prompts.
You can also test this. Put bad_hands into positives instead of negatives, see if it actually gives worse hands on average. For me it does only as long as I do nothing else with the hands, if I add waving, grabbing [item], clenched fist, etc it gives good hands even with "bad hands" in the positive prompt. Same thing with artist tags negating jpeg artifacts, film grain, sepia, etc.
friendly reminder that anyone actually advocating for "no negatives" is a blind retard who could look at a red painted barn and claim it's a yellow schoolbus.
>>8894911 Honestly, I was hoping there would be some sort of methodology/criteria that would allow for the creation of a mathematically "ultimate" neg - one that includes as many undesirable things with as little of a creativity/flexibility penalty as possible. Obviously you can't actually evaluate any of this quantitively, but by keeping in mind what potential candidates for neg tags do and what their side effects could be, perhaps it would be possible to at least approach that "ideal" instead of randomly throwing shit at a wall with no rhyme or reason.
>>8894919 You just hate real world too much, that's why you put ye-pop in negatives. You want the creatures you generate to look as flat shaded and simplistic as possible, because if they don't they'll start reminding you of an outside world and all the bullying you've endured, probably from girls too. You HAVE to use the negatives, otherwise your mind will break. But I don't have this problem.
>>8894887 >>8894919 this person is vantablack >>8894921 "negative/positive prompt" construct is part of cfg, aka classifier-free guidance, which is always used to generate images (unless it's distilled or whatever). and there's really no "negative" prompt, this is just the retarded way A1111 decided to call "unconditional" prompt.
for the positive outputs, the model MUST be trained on well-captioned images. for the uncond outputs the model MUST be trained to denoise whatever it can see in the given image, which is called unconditional training. to achieve this, the model MUST be trained to denoise images with some sort of anchor braindead prompt, be it an empty prompt or just a zero embedding. this is why it's called "classifier-free" guidance, because usually for the task of class-based image generation it's easy to train a separate guidance model. with prompt-based image generation, you can't really train a guidance model and you have to resort to weird tricks.
the point of cfg is to pull the output distribution into something specific from the sea of dogshit, emphasize your specific prompt. cfg does that by calling the model twice: once with the positive prompt, once with the "negative" prompt. then it combines outputs according to this rule: uncond_output + cfg_scale * (positive_output - uncond_output)
the bottom line is YOUR BASE MODEL IS FUCKING LOBOTOMIZED if the outputs with no negative prompt look like ass. your "mathematically ultimate" neg is such a neg that was used to train the model unconditionally.
>>8894951 >the bottom line is YOUR BASE MODEL IS FUCKING LOBOTOMIZED if the outputs with no negative prompt look like ass. I mean I can't really argue with that after such a write-up, but it's not like I can choose or meaningfully modify the models I'm using. All I can do is change the pos/uncond sloppa on the lobotomized model whose outputs with no negative prompt look like ass since that's the best I've got, so saying "this model is bad, use something else" is just unproductive.
>>8894962 >but it's not like I can choose or meaningfully modify the models I'm using you should start by spreading the word. there's still a chance it would be heard and the next model we'll get is not going to be fucked up. >All I can do is change the pos/uncond sloppa on the lobotomized model whose outputs with no negative prompt look like ass since that's the best I've got actually, if you really think about it, you can try to fix it with a lora. normally, captions are dropped like 10% of the time, meaning for a 10M dataset you'd only need to train on roughly 1M images without any captions for 1 epoch. but that alone would likely collapse the model, so you'd have to introduce some additional preserving loss.
>>8894975 >you should start by spreading the word. there's still a chance it would be heard and the next model we'll get is not going to be fucked up. I'll keep it in mind going forward. I had my suspicions something wasn't trained quite right after that NAI vs Anima style comparison from a few days ago. >actually, if you really think about it, you can try to fix it with a lora. normally, captions are dropped like 10% of the time, meaning for a 10M dataset you'd only need to train on roughly 1M images without any captions for 1 epoch. but that alone would likely collapse the model, so you'd have to introduce some additional preserving loss. I can't imagine that being possible on consumer hardware, not to mention gathering such a gargantuan dataset. That's why I'm keeping it humble and just trying to make the best of what I've got by only tweaking the prompts.
>>8894981 >right after that NAI vs Anima style comparison from a few days ago this likely has little to do with cfg, it's just dataset curation and processing. furthermore, nai uses its own negative prompt. >not to mention gathering such a gargantuan dataset you can always start small, you can get visible effects with as little as 10M images