>>8881951 People explored all the NL shit when anima was fresh, like all those cuck phone holding gens. But ultimately most people just want simple peepee in vagoo.
@grok come up with creative hentai positions that I can prompt on a 0.8B TE and will make me get lots of (You)s on /hgg/ home of the most creative local genners of the internet, make no mistakes
>>8881963 --xformers --flash --sage --cuda-malloc --cuda-stream --pin-shared-memory --lowvram --lowvram might seem like an odd choice but in my testing it doesn't affect the speed whatsoever, it just makes OOM crashes/slowdowns less likely to occur by unloading parts of the model that aren't currently being used in more cases instead of keeping them in memory, made a particularly big difference when Hires. fixing
Why is anima so horny? I literally don't prompt anything but quality tags and pretty tame artists but it always tends to show ass, panties and tits. Don't think I've ever seen a model this nsfw-inclined
>>8881970 Thanks anon, that shaved off a couple of seconds from my gens. I know there's some concerns with these settings affecting gen quality but I don't know how much of that is bullshit, either way in my case with my shitty 8gb of VRAM I don't think I have the luxury to care
>>8881978 None of those settings affect quality to a noticeable degree the way something like fp16 accumulation does. Also forgot to mention Torch Compile, set it to "dynamic" for a further boost in speed at no quality cost, and since you have VRAM usage to worry about, I'd also use Never OOM for VAE at all times and for UNet every time you want to go beyond 1216x832.
>>8881951 it may be able to understand them but ultimately its still a 2b model, when a model can do anything anyone wishes for without mistakes, then you will see much more creativity
>>8881981 It took me an embarassing amount of time to find those options in the UI but yeah the dynamic really does make it gen significantly faster. Based anon spoonfeeding a retard, I'll pay it back by posting gens soon
>>8881951 Stuff can be made. But it takes lots of wrangling not to get weird anatomy. Also one needs to study cinematography and art theory to know all the words and phrases they use to describe mages.
>>8882046 Comfy here so I'm just guessing. But it's not a "proper" controlnet, it's kohya's LLLite gimmick so you need support for it specifically. He claims it's supported in sd-webui-controlnet, but based on the log you're using sd_forge_controlnet.
>>8882048 Pulled 2 days ago, pulled again now to no avail. Can you catbox a screenshot of your img2img tab set up for inpaint CN? Maybe I'm missing or using something really stupid. >>8882050 There's been a commit specifically for Anima LLLite 5 days ago.
Anima may not have a good controlnet or inpaint but this one is just fucking good on base anima: https://civitai.red/models/2587631/nude-filter-or-simple-edit
For absolutely no reason, it can nude something without changing the style or even hair infront of the body, you can even slap a lora you baked and it'll apply that type of body whilst keeping the style of the image.
>>8882067 >Anima may not have a good controlnet or inpaint The inpaint CN is actually good. Anytest is a bit inaccurate but still alright. The Cosmos ones we got before actually suck, and have style bias from the lora you have to use them with. This seems to be another one of those, and not even for Anima specifically.
>>8882081 I always had a problem with inpainting, no matter what i'll keep having seams, body issues or alternating the background even with noob. I even asked for a workflow here and attempted via forge it still has skill issues to me. this one is good at one thing and it doesn't change the background or make those feathering type of blur.
>>8882063 >had to set Preprocessor to None I fucking swear with this shit... How was I supposed to know? It even sets it to something else automatically when I select "Inpaint". Thank you anon.
https://litter.catbox.moe/12v2496gq493nd6k.png >>8882100 I had issues with the preprocessors, but it was just adding random noise instead of inpainting without printing errors. Glad you managed to solve it.
>>8882100 Does Forge allow you to see the output of those preprocessors, before it's passed to CN? The classic one blacks out the inpaint area so that inpaint CN won't see it. This one relies on generic CN masking instead.
>>8882113 It's not that the preprocessor does something wrong, it's that it does nothing at all as if it's being skipped entierly. If I set it to the default of inpaint_global_harmonious, make a failgen, then disable CN and redo the inpaint with the same seed, the output is completely identical.
>>8882117 Pretty certain those middle steps never leave RAM, the only outputs that aren't obfuscated are the image at the last step and the low-quality live previews.
>>8882067 wtf. how the hell does this thing work? looking at the preview it only denoises part of the image. how does it know how to do that?? agi?????
>>8882121 Well most of them are standalone tools frankensteined into webUI, like the pose estimator, depth map, canny edge detect, etc. They process the reference image, spit something out and pass it to the controlnet. So I assumed there would be some way to see it.
>New model >Has NLP >Updated dataset means more artists work, especially ones with low image count >Base gens look way better >Still pretty fast Yeah wow, this is grea- >NO ARTIST MIXING
>>8882154 I certainly don't miss it, I've never used more than 1 artist at a time anyway. I want more of the styles I know and love from VNs, doujins and gacha, not some mystery meat slop I have to waste hours getting to through mind-numbing trial and error.
>>8882192 Did you spell it wrong, or did you slop in so many loras and quality tags you can't see the style through them? Seems to work fine for me with "@murata range"
>>8882063 >inpainting v1 Granted I haven't fucked with these yet, there is a v2. Are you aware or found v1 to be better? Also, can you link the git for the aspect buttons? I pulled one after moving to Neo and it fucked my install up.
>>8882204 >Granted I haven't fucked with these yet, there is a v2. Are you aware or found v1 to be better? I just wasn't aware >Also, can you link the git for the aspect buttons? I pulled one after moving to Neo and it fucked my install up. https://github.com/altoiddealer/--sd-webui-ar-plusplus
>>8882203 Was the pic at >>8882192 not an example of the failed attempt Oh, the other murata. His booru art is all over the place, he doesn't even have a style. To be expected
>>8882278 >why wont @ tags work @artist works. Even ones with low data on the boorus. Are you on a slop mix model or using something like euler a that smooths details out? >save settings on controlnet/inpaint, even this shift meter Settings -> Presets -> Anima
>>8882063 >>8882100 Thanks, anons. Through both of you I got this new inpaint model running. Works really damn well. First time I've ever not had to use softinpainting.
>>8882289 kys nigger, patreon grifters can take 1000 nigger dicks in every orfice. It just pisses me off when I see a decent gen and I want to check it out to gen it with different characters so I can shoot a load but some faggot thinks they're Michelangelo for writing some prompt and prunes their metadata. And then I have to BEG for some fucking catbox and I'll be lucky if I get it, fuuuuck you niggers
>>8882154 i havent tried it yet but why wont shit like (@artist1:0.5)(@artist2:0.5) work (yes anima supports prompt weighting i just dont know how weights less than 1 behave)? i guess itll just output garbage? even then, why not go full classic a1111 syntax and use [@artist1|@artist2], this should be pretty doable in comfyui even without syntax support. even if the llm doesnt like it this should give results
>>8882301 >why wont shit like (@artist1:0.5)(@artist2:0.5) work The opinion seems to be that sort of "mixing" was a happy incident of CLIP. Not intended but capable due to the bleeding effect. >classic a1111 syntax and use [@artist1|@artist2] This is what you use to mix on anima. It's more complicated only because what artist you use to start or end with and at which step has to now be taken into consideration. It's doable but it takes a bit more testing than just tossing 4 artists at CLIP at random weights and hoping for the best.
== Prompt Workshop/Booru Grabber == - 6 input boxes in Anima tag order (quality/meta, subject, character, series, artist, general) - Build Prompt consolidates + auto-deduplicates into correct order - Send to txt2img appends to positive prompt (never overwrites) - Per-category Send buttons for partial sends - Auto-formats: lowercase, spaces not underscores, @ on artists, score_N preserved - Pulls random Danbooru or Gelbooru posts, normalizes tags to Anima format - Search by tags (multi-word, combinable with rating filter) - Rating filter: All / Safe / Sensitive / Questionable / Explicit - Tag blacklist + category blacklist (e.g. skip artist tags entirely) - Grab by post ID, Open Post in New Tab
== Anima Resolutions == - Resolution picker: inline dropdown + 8 quick buttons below txt2img seed row. 36 presets, all multiples of 64, W+H ≤ 2176, covers square/landscape/portrait/ultra-wide/ultra-tall, 640px minimum - Auto-scale for img2img: when sending to img2img or inpaint, sets Resize-by scale to land post-upscale in Anima's 3072–3264 training bucket range. Reads image metadata, not sliders - Randomize resolution: checkbox that picks a random preset each Generate click (runs before_process, one roll per batch) - Session prompt history: 100-entry LRU ring of your positive prompts. Dropdown with ~160 char preview, select to refill prompt. Click Refresh to update. Clears on webui restart
Also recommend er sde with beta and 4-6 cfg. 40 steps is kino. Do img2img upscaling with 0.6 denoise (25 steps if your step count is 40 in txt2img) and DONT use a img2img upscaler at all - extension auto fills the correct scale. Make sure to turn on "Send the Parameters in the infotext instead of the UI fields when using the "Send to" buttons (e.g. send the result of Wildcards instead of the syntax) (requires Reload UI)" in webui settings so the same seed is used in i2i. weeee
>>8882383 because we live in a modern society. plump bodies are no longer signs of fertility or wealth, but are instead signs of poverty and mental illness.
Hello is this the best way to train Anima lora on Windows? https://github.com/gazingstars123/Anima-Standalone-Trainer Which tagger should I use and what settings? Never trained a lora before Thanks
>>8882466 >is this the best way to train Anima lora on Windows? no. it's good for beginners i guess. use this https://github.com/67372a/LoRA_Easy_Training_Scripts or just use sd-scripts if you are familiar with terminal environment. >Which tagger should I use most bakers use wd-eva02-large-tagger-v3 for illustrious/anima. for ui, i recommend taggui or Jelosus2/DatasetEditor.
>>8882503 if a person really wanted it they'd come to my house, push me to the wall, grab my neck and DEMAND it from me with a husky growl that sends a shiver down my spine
>>8882498 >because there are too many dishonest asks this only matters if you know and recognize what you're posting is garbage. in which case why are you posting it?
>the inpaint trick from pre-character reference NAI days works with the Anima CN, just barely Imagine if we just got an actual edit model or a non-abysmal dogshit IP-Adapter. Instead we're about to get... turboslop (for a model that's already acceptably fast).
>>8882516 It's not a cope, they do work, better than anything else if you just want raw quality and know what you're doing even, the problem is that they're impractically bothersome to make, if enough reference material to make one exists at all. One-shot solutions (preferably architecture-agnostic ones) that are 90% as good but take 0.000001% of the time and effort is what local should be focusing on right now, it's inexcusable when even NAI got it figured out, let alone safetycensorslopped SOTA.
>>8882536 It's not just about speed, it's about the dataset too. Sometimes you just don't HAVE 30 images of a character to bake a LoRA, as in there doesn't exist that many on the whole wide internet because the character is the literalest of whos, then what do you do? To my knowledge 1-image LoRAs are impossible to make work.
>>8882538 nta but i have baked some usable character loras with just 1 image. I used nano banana and chatgpt images (and novelai...) to generate some more images for training. of course the quality is much worse than using real images but it's still usable
>>8882538 >what do you do Could you use an edit model to generate some more images? But then you end up training on synthetic data, so maybe it's not such a good idea.
>>8882537 It's model-agnostic, happens purely on the UI level. Like regional prompter. Works on any model that takes a prompt and runs in steps, even vidgen.
>>8882559 >But then you end up training on synthetic data, so maybe it's not such a good idea. Is that such a bad idea? With the quality of the output and manual verification, it shouldn't be such a problem.
>>8882562 There's something about feeding the model AI outputs that aggravates all the common imagegen issues. Makes missing limbs, broken object permanence, bad hands, etc all just a little bit more likely. And it adds up the more you iterate.
>>8882555 A waste of time most likely. >>8882533 Why? Turbo LoRA works just as good in terms of speed, but it doesn't sacrifice artist knowledge, unless something has changed significantly with RDBT recently. Guy can't even figure out how he wants to train the model and even typed the wrong version name. He made so many versions trying to combat different problems, problems that Turbo LoRA didn't have at all and was just 1 version. Maybe the newest version of 0.35 works as good as Turbo LoRA Also why do you specifically need a LoRA? You could download the finetune RDBT if you really wanted it.
>>8882609 This shouldn't work, different use case.
>>8882596 Main issue with these edits, even when done by real artists, is they don't account for the boobs being compressed in clothing. The size and shape should be different when naked.
>>8882625 Yes and you can inpaint a nude girl in the area, but not the same girl. Because you need to also reference the inside of the inpaint area for that.
>>8882633 >lr 1e-4, dim/alpha 32/32, batch size 2, came + rex isn't the lr a little too high? i use lr = 4e-5 with alpha = half dim and it always feel like the lora starts overfitting at around 1000 steps
>>8882647 >exactly $0 i am proud of you. never spend money on virtual funkopops that canonically love big black men, not you. buy your amigo de madre something nice instead
i've been using swinv2-tagger-v3 at 0.45 weights for a long time, is eva 03 better or something new for auto taggings? I still manual review of course.
>>8882661 shit, its V2-large-autotagger-"v3" my bad.
although I did see some new auto taggers but for some reason its in a safetensor file. https://huggingface.co/animetimm/convnextv2_huge.dbv4-full No idea how to use this on tagging softwares.
>>8882665 >animetimm/convnextv2_huge.dbv4-full oh yeah i'm using this tagger with this comfy workflow https://files.catbox.moe/jaqhzl.json i'm not sure if it's better than eva02, but it definitely knows newer characters.
>>8882668 .toml holds settings for easy-training-script you can open it up in a text editor and translate into kohya's command line arguments, if you wish
>>8882677 MultiDiffusion was a cope workaround to get a ghetto Hires. fix back when the early previews couldn't go past 1024x1024 at all, you don't need it anymore. But if you insist, you set it to Mixture and Diffusers, enable it, paste your image into the img2img tab, set the Resize to dimensions to the target resolution (Resize by does NOT work) and set the MultiDiffusion tile width and height to the input resolution.
>>8882666 this is taking long to tag, after 10 mins it kinda just finished without making a .txt file, excluding downloading the model. also wonder if it can replace the () with the usual /(\)
>>8882712 because its boring as fuck you have the most powerful artistic tool ever conceived and there are people who just want to make anime screencap images that look flat and shitty its so retarded anime only looks like anime because it is limited by humans, now we can make anime more advanced as it should be and cast off the shackles of "creatives"
>>8882626 >>8882633 >>8882671 How does something like this compare https://github.com/ThetaCursed/Anima-TrainFlow to easy training script? But I guess when you don't know what any of the numbers do then it doesn't matter. When will we have a proper training guide that teaches the basics? Not something like the rentry in OP that covers useless shit like how a model works, what models exist and that entire glossary shit which you'll never touch when actually training. Someone needs to make a proper rentry that covers training Anima in Easy Training Scripts, explaining what to set all the settings to and a brief explanation why. Not teach someone useless shit like what a LoRA or a checkpoint is, if you don't already know, you shouldn't be training. It's like Blender, can do a million things, but if you want just one thing, it fucking sucks unless you've already mastered blender.
>>8882737 It won't happen because the moment someone publishes a guide someone will come out and call the guide bullshit and then there will just be shit slinging. The best way to learn how to train a lora is to literally just train and then check your results and then repeat until you are happy, which may or may not drive you insane. There is honestly no point in doing anything for this place anymore.
>>8882666 >>8882704 I figured it out, it just feels like the same as eva02.
>>8882737 I honestly don't know how ezscipts works, even loras in general hell I even had a problem long ago asking why it won't do the start training but somehow someone will helped here and if it works its kinda a set and forget. and any of the technical terms here I still don't know. those rentry and articles in op are useless to me too, I just believe in one guy's tomls here who I always ask questions even ever since using the old kohya's ass one.
>>8882742 Playing trial and error is just retarded. There are people already more than capable of making a proper guide, like the one who made the rentry in the OP. He just needed to not rant like a fucking schizo. Sure there will always be some people who fling shit, but I believe a lot of genuine posters would appreciate a proper guide. >>8882745 >set and forget Sure, but would be nice to understand some rough basics about what the settings does and documented by someone who actually know what they're doing roughly what settings should be used, rather than a settings file given by a literal random that maybe works "good enough", but since you don't know any better, it could maybe be 10x better because you just trusted that it's good. We need proper documentation without schizo babble.
>>8882714 girls in anime style are the most beautiful of all possible variations and forms of art, nothing better has been invented yet also your experimental shit won't be anime by definition, because anime has own defined style get the fuck off this thread, by the way
>>8882756 NTA. Last week the artsy-fartsy community in reddit and twitter had a bit of a flame war because some chinese dude could draw hyperrealist pictures. Some of them called it not "real" art. I just don't know anymore.
>>8882737 >>8882747 >rentry in the OP tbqh i sort of abandoned it because literally nobody cared. i was asking anons what's wrong with it, what should i add or change, but i only got very few actual on-topic replies. like, what's the point of writing if i am not getting any feedback.
if anyone wants to take over it, feel free to do so. >>8882737 >useless shit like how a model works that's only useless to a point. once you get the basics done, you'd want something more.
>>8882766 >that's only useless to a point It's not useless to a point. It's completely useless. The guide is for training a LoRA, you don't need to know how a model works for that, just what each setting does and what is good.
I use Forge Classic and I searched the webui settings and the backend/args.py for "torch" to set Torch Compile to "dynamic", but I'm not finding out where to do that, or with what flag.
I saw you can do it easily in a ComfyUI Torch Compile node. Is it just a ComfyUI thing, or even just a Forge Neo thing that I can't use on Forge Classic? Or am I just missing where to set it?
>>8882737 here's the inconvenient truth. a majority of lora training is vibes. and a lot of the time the "answer" for why settings are what they are in configs people post end up being one of three things. 1. I kept adjusting them wildly in all directions and feel like this works(sort of technically good reason) 2. it got set at some point and it hasn't given me reason to think it's causing a detriment and hasn't exploded so I haven't touched it since (mediocre reason) 3. vibes. (lmao) 4. I think I read about doing it this way somewhere maybe? iunno. (bad reason)
there just usually isn't all that much thought or reasoning put into most of it. Some decisions are grounded in repeated iteration but even that is kind of worthless if something is wrong somewhere and skewing results, but other reasons can be boiled down to just not really touching them past initially setting things up. and most configs really are just "adjust LR up or down pending results and try again." They're not very complicated.
and at the end of the day config shit is probably the easiest part after you have something that works. The hard part is if you need a different config for different types of loras(read: style, character and concept) and what dataset prep should look like between those, which you can get like 50 different answers for. And then the really hard part is understanding the really stupid things you can do to brute force into either category, which you'll only really understand once you've trained a ton of loras and folded them over a gorillion times while doing a lot of stupid things as experimentation.
>>8882554 As I am retarded, is this more akin to the dmd2 or cfg version? Because cfg at low weight is my absolutest black negro. The one for preview 3 still works well enough but new thing.
>>8882783 you keep getting told you're retarded because you are. multimodals will still need style loras(probably significantly more than regular models, unless you're content with basic bitch GPT tier "anime") and more importantly than that they'll need concept loras for whatever deranged shit you're trying to do that you can't just do with a regular, normal model.
>>8882780 >And then the really hard part is understanding the really stupid things you can do to brute force into either category imo it's the easiest part. once you'd realized that the model is a retarded dog and it will do whatever you train it to do, you'd reached the enlightenment.
>>8882787 The whole entire point of models like that is that you can show them one reference image, tell them what to do with it and they'll just do it. The reason you think they would need style LoRAs is because none of the models capable of such a workflow have been trained on the 'booru. No shit they can't replicate specific anime styles when nothing in their training data differentiates them.
>>8882791 Anima does well enough with low-image artists who have consistent artstyles, the real actual problem are oldfags with lots of images whose styles evolved over time. It just can't pick out the style of any given period so you get a bad-looking averaging of them all.
>>8882807 are you trying to imply your slop looks anything remotely similar to yoshitaka amano? but yes, most of yoshitaka amano's work is pretty japanese, albeit very high fantasy type japanese. however it's pretty much the opposite of erotic and would look like absolute garbage for anything actually intended to be porn.
>>8882814 >however it's pretty much the opposite of erotic and would look like absolute garbage for anything actually intended to be porn So is every single noodlekek style btw
>>8882817 seems about to be expected tbdesu. this is pretty much the type of stuff he excels at https://danbooru.donmai.us/posts/71232 And look, there's actually an explicit tiddy out. But you don't really notice or care because it's just kind of naturally there. And that's fine, nudity doesn't have to be erotic in art. Hell, in some sense it's actually harder to have non-erotic nudity. I think he has a problem though in that he's cursed to the point that he can't do anything erotic even if he wanted to. he'd probably be absolutely killer at japanese high fantasy'ing nihongo folklore/youkai stuff though.
>>8882812 If you're gonna gen pixel art, at least use the extension that downscales and rescales back up to fix those horrendous pixels so a pixel is actually a pixel. It'll also save you a significant amount of space, just look at picrel, 97.36% smaller and clean pixels instead of blurry shit. Ideally also gen in a resolution that can divided by the amount of pixels it's trying to gen for that artist, like in this case 40, which you can't divide 1024 with. Closest to 40 would be either 640x640 or 1280x1280 You can use an extension like this: https://github.com/mrreplicart/sd-webui-pixelart Or just do it manually.
Is it just me or using a mask with controlnet on forge neo does absolutely nothing? I also tried using the brush directly on the original CN image but I don't see any difference
>>8882851 oh nyo. That was my reasoning for a few extensions when I moved from reforge to NEO and it fucked my install up. Might just make a dummy directory to test.
>>8882853 If you want, I can just give you my setup that I use to fix pixel art, it's just a batch script that relies on Magick, you put image into the folder, run the script and enter the downscale factor and it does the rest. It's neater to just have it as an extension in the webui if it works though, saves you from moving files around manually.
>>8882854 Yeah that'd be cool just in case. It's not something I dabble in too often but it'd be nice to have the option when I want actually try pixel art.
whats wrong with forge neo, adding an @artist tag wont work, but once I add any lora and then an @artist tag it suddenly works with blending the artist. the non lora ones seem to look low quality sketches too.
Updated the extensions I posted yesterday with my buddy Claude, someone might find them useful. Also thanks to the anon who suggested going to simple instead of beta schedule, it's so much better for everything and makes sense russell used it in all his examples.
Found my first prompt the other day. Happy that I can easily make it work today.
>>8883043 To be absolutely fair, I have managed to get very close. However, I often need to lower the strength, to try to change the style, but then she loses the dozens of little details she has. It's a mixed bag of dicks.
huh. trying to gen zhao \(zenless zone zero\) on anima and it just doesn't know the character. like, at all. period. Even trying to add in a bunch of supporting tags and not using loras or anything, it just doesn't know her. but the character should be old enough to show up in the dataset in more than large enough numbers so I wonder what the fuck happened to cause this
>>8883093 >>8883094 oh weird I would've thought it was a bit later than that though I guess maybe I should actually read to find that sort of thing out.
>>8883101 there is a zhao lora if you want https://civitai.red/models/2586211/zenless-zone-zero-all-female-characters but there is no scat lora for anima yet if you are going to gen that
>>8883150 I'm happy to do that, why don't people post in /hdg/ anymore? I left when people couldn't stop complaining about highlights, and when I came back we had two threads and this was less inactive.
>>8883220 someone taught them how to gatekeep incorrectly, as a joke. as a result anyone who gave a shit left ages ago and all that's left are webgenners, a few of the lowest common denominator local users and, the centerpiece, a small group of people who don't even gen but feel the need to control the thread and shout down at anyone who wants it to improve.
>>8883218 >a bunch of batch files It's literally just one lmao. Also I made it ages ago, long before you retards even knew about ShitUI. You're also comparing apples to oranges, writing something vs using something existing. How long would it take you to actually write that ComfyUI extension which most likely came into existance long after my setup? And like I said, not everyone uses ShitUI, just like me and the guy I was talking to. I wasn't talking to you now was I? My setup is already made, it's not any more difficult to download than setting up an extension in ShitUI and it doesn't require having ShitUI installed at all. Why didn't you link your "superior" ShitUI extension in the first place if it's so great? It's better to use an extension as you're genning, if you can in whatever UI you use, but if you haven't used it or for some reason need to do it on existing images, then my setup is better, can be used instantly as opposed to ShitUI that takes time to boot. It is literally the simplest and most lightweight way to rescale images.
>The Anima Model is licensed by CircleStone Labs LLC. Copyright CircleStone Labs LLC. IN NO EVENT SHALL CIRCLESTONE LABS LLC BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL. >Built on NVIDIA Cosmos Wtf is cprussell cooking
>>8883288 Do you want something simple that just works? If so then Forge Neo. Do you want something overly complex that accomplishes exactly the same but with 100 steps instead of 1, making you wish you killed yourself instead? If so then SpaghettiUI.
anyone got a inpainting workflow for anima that behaves like forge neos? I don't wanna keep opening that just for the sole purpose of inpainting. the one given in llite is alright but not as good as forge's
>>8883235 chill bro, you're literally writing something to use something existing >which most likely came into existance long after my setup? imagemagick extension is over 2 years old and image resize nodes existed since comfy inception lmao >need to do it on existing images, then my setup is better, can be used instantly as opposed to ShitUI that takes time to boot. with comfy alone you can do whatever the fuck you want to any images, not just the ones you generate. also, >implying you're not running comfyui behind gradio ui (forge) all the time anyway
>And like I said, not everyone uses ShitUI, just like me and the guy I was talking to. I wasn't talking to you now was I? i just pointed out an easy way, the way retards like you are malding over the other tribe is honestly still pretty hilarious
>>8883304 He's not wrong, though while you resist killing yourself you end up leaning some basic of how this shit works under the hood. Afterwards they're roughly the same, once you have your own workflow set up for all the common tasks.
>unnnf hnnnngg oohhh~ bra-BRRAAAP bRRRAAAAP BR-plop-ppppffffffffttttt.. pft little brown bros still being produced at the maximum capacity sir :thumbsup:
>>8883331 yeah this is the old maskdetailer, this one always give me problems even in noob. I used the same setting as the rustle looking pic thing in forge neo: >>8882063
This one makes it good, its near seamless, no outer image manipulation and no changing that odd "bbox" setting as it's a 1:1 sizing. it's this setting I want into comfy.
Now that anima v1 is finally out, do we have any better way of upscaling than using multidiff? Also why the hell does the anima lineart controlnet require black lines on a white background instead of the usual white on black?
>>8883321 here's mine: https://files.catbox.moe/fkyf70.png as for how to control padding, it tries to fit all your masked areas into the image context, so just put a little dot or line away from your main mask
>>8883417 thanks for the workflow, it kinda works well enough. I noticed torchcompile did help with the image gen, shaving ~3 seconds. you got any other tricks on speeds? I think on my bat file it has: --windows-standalone-build --use-sage-attention --cuda-malloc --disable-api-nodes
>>8883417 >>8883412 shit I also forgot to ask, what's RDBT I saw it in your workflow but I did remove it during the inpaint testings, and more importantly the "tiny dot out of nowhere" trick did a huge help.
>>8883428 none that i've tried. i've heard people use easycache with anime >>8883431 >RDBT it's just a lora a guy here recommended. tried it and ended up really liking it, better composition and stuff >"tiny dot out of nowhere" trick did a huge help. yeah that trick works for every ui, completely removes the need to mess with that stupid fucking padding slider
>>8883566 It would be dead regardless. Also with two threads it's more often that case that at least one is not at bump limit so some newfag can occasionally stumble upon it. With one thread it's weeks of being on some page 69.
>>>/g/108863598 they will shrink me down and force me to climb up into her ear. the rest of my sorry brown life will be spent shoveling ear wax and cutting down ear hair
>>8883325 My setup is literally from 2023, before you or anyone else knew about SpaghettiUI lmao >>implying you're not running comfyui behind gradio ui (forge) all the time anyway Never implied that, but my setup doesn't require launching anything like that which takes actual time to launch. >with comfy alone you can do whatever the fuck you want to any images, not just the ones you generate. also, Sure, but why use bloated tool that takes time to launch if you don't need it. It's like using Photoshop when all you need is to draw a black square onto something which can be done much faster in MS Paint. Just because a tool has a lot more capabilities, doesn't mean it's better for every user. >retards like you Says the guy who has zero reading comprehension lmao. Take your meds schizo
>>8883623 Yeah. It's pretty easy. You get dogpiled and harassed by anima shills if you use any other model; and nobody's actually using it after the honeymoon period. People are back on pony and illustrious, and they just avoid posting here because they don't want to be attacked by the shills.
>>8883625 Stunning & brave post, we need to speak up against the Anima shills like we did against the noob vpred shills and reclaim this space for real creatives.
>>8883623 It's all your fault for declaring that you are only here to shitpost and engage in dishonest practices and for encouraging others to do the same
>>8883709 It comes alive somewhat during burger hours, but yeah. In my case I just don't feel like genning anything. Already did all I ever wanted with noob, anima just makes the same stuff way easier.
>>8883719 DiT models take square time to generate depending on resolution so 2MP gen is 4 times slower than 1MP gen. Maybe with 5090 it's fine, but even then I'd still prefer speed.
>>8883726 >it can, even if it isn't stable You just answered it. T5 is the middle of the road between CLIP and LLMs so it's halfway between the two when it comes to artist mixing - not entirely consistent like on CLIP but not entirely gone like with LLMs.
>>8883709 Because they excluded E621 knowledge and while it beats illustrious base model, it's not all that insane compared to a finetune of illustrious. But now that it's done I'm sure someone will properly finetune it.
>>8883737 It IS slower. I usually need only 20% denoise for upscaling anima so that's 1 + 0.2 * 4 = 1.8x compared to 4x if I just generate at 2MP. And then of course the fact that I upscale maybe 5% of the images because I have standards...
>>8883726 llm embeddings provide very little actual visual cues to the diffusion model unlike clip embeds. like, llm embeds say "draw this in the style of X" while sdxl clip embeds say "if you want to draw this line in the style of X, draw it here, make it this thick, choose this color, etc"
>people who gimped themselves by trying to focus ONLY on artist tags for mixing while ignoring loras are being gatekept from style mixing on anima lmao
>>8883747 If it takes as long for you to do 25 steps in img2img at 1.0 denoise as it does 0.01 denoise, then you're doing something horribly wrong. Lower denoise is significantly faster. 0.5 denoise is twice as fast as 1.0 denoise
>>8883763 No, you get them from higher denoise, at the risk of losing original structure the higher you go. There's no reason to go as high as 25 steps in img2img, unless of course you use batshit insane denoise of higher than 0.3
>>8883767 >the dmdx lora looks exactly like the official turbo lora Is it faster at least (i.e. needs less steps for the same image), or else why would you use it over the official one?
>>8883768 No I already use 12 steps at 0.2-0.3 denoise and it works really good, increasing steps doesn't make any significant difference compared to raising the denoise. Try using your brain at 1.0 instead of 0.0
asking the author of the BLACKEDRAW-anima-test-2 lora to please provide trigger tag or usage instructions, it doesn't seem to stop my anime girls from being anime
>>8883772 for anima at the moment, if you're upscaling, .5-.6 seems fine, varies image to image though. Problem is it still has really bad dimple/noise type artifacts. Once we get controlnet tile it's .7-.8 denoise or it's not even worth bothering with.
>>8883776 You're retarded if you're trying to upscale something with higher than 0.3. A denoise that high should only be used for inpainting, when you genuinely want structural changes. Otherwise you need it fairly low.
>>8883769 The outputs look exactly like the official lora, with the same step count and seed. This lora was warm-started from the official lora, and it didn't learn shit. It doesn't change the ouputs at all
>>8883779 >for anima at the moment, if you're upscaling, .5-.6 seems fine, varies image to image though. Problem is it still has really bad dimple/noise type artifacts. yeah, that's also my main complain about anima
BTW if your image loses texture too much and comes out not as detailed as you would expect after upscaling with Anima, you don't need more denoise or more steps. You need to give it slightly more noise than schedule requires (like NAI extra noise slider). No I'm not providing a/b comparisons or proofs. Feel free to ignore this and keep using 0.6 denoise, I shouldn't be sharing my godlike secret techniques with plebs anyway but I'm feeling generous today.
>>8883779 If you just want a carbon copy controlnet for upscaling, you can use the inpaint one. It needs a mask, but you can give it one pixel somewhere in the corner which basically tells it "the rest stays".
>>8883853 what's the obsession with multiple penises in the same hole? not only is dicks touching mega gay, they are lathering their penises with each other's cum/precum and smegma
>>8883858 If you can train at 1024px then don't train at 512px. Training at 512px is fine for simple art styles and characters, but 1024px can capture more detail
Noob had the issue of having the man peeing in the girl's asshole every time you prompted for anal+pee, but anima has the exact opposite problem https://files.catbox.moe/sxoxdx.png
>>8883903 Oh it's inpainted, OK. But why is man's pee crystal clear? Girl's pee should be clear, man's pee should be yellow and nasty and with yellow clouds of death coming off of it!!
If anyone has problems like I did with viewing catboxes, try this in a powershell run as admin Set-DnsClientServerAddress -InterfaceAlias "Ethernet" -ServerAddresses 1.1.1.1,1.0.0.1 ipconfig /flushdns
Replace "Ethernet" with actual interface name which is found by doing Get-DnsClientServerAddress in powershell
>>8883833 after raping gemma-chan i can conclude that as long as she's in good condition, very careful and you're not a biohazard, she can even manage sex and give birth to children and everything associated with it.
>>8883924 >made for humans Microshit controlpanel is not made for humans, finding shit in there is a pain. Pressing Windows button and then typing in powershell to copy paste a couple lines is much easier and faster. Claiming otherwise, is bullshit and you know it.
>>8883932 There is literally nothing wrong with powershell. Powershell lets you do things instantly, doesn't require you to search for shit unlike controlpanel. Don't see how this relates in any way to Linux. Rent free I guess lmao.
>>8883938 verbose syntax is gay, reliance on dotnet is extremely cringe and very gay, execution policy and permission bullshit is very cringe and super gay. command prompt is outdated but at least it's not gay. >>8883941 yeah in powershell instead of the gui panel you search for the right command kekmao
>>8883947 >instead of the gui panel you search for the right command Searching for the right command is much easier than trying to find something hidden beneath 100 layers in the controlpanel that you have to search for anyways lmao
>>8883973 >need to find thing in a gui >click around till you find it >need to find thing in cli >MUST google it because none of your options are laid before your eyes, if you don't know a command exists you have no way of naturally discovering it clitards are so out of touch with reality
>>8883977 comfytards love pretending they have exoteric sysadmin knowledge because they tortured themselves into learning the most obtuse node based interface concocted by a human being
>>8883978 you turn your pc into a linux server and run whatever gui in there, then access it through the vpn over internet whenever you want obvioiusly
>>8883976 Yeah because it's such a pain to google something right? Are you allergic to searching or something? Chances are you're gonna have to search shit anyways the first time you do it because you don't know wtf you're doing anyways. Let alone what numbers to put in, even if it's just 1.1.1.1 lmao You have now searched and found the command, if you ever need it again, just store it somewhere easily accessible. >>8883977 If only it was that simple, it's more like rightclick taskbar network icon, go into network settings, change adapter settings, rightclick ethernet, properties, find IPV4 shit in the list, hit properties then finally after all that bullshit, change to 1.1.1.1 lmao
here's how you change your dns if you use dnscrypt-proxy on linux you open up the config /etc/dnscrypt-proxy/dnscrypt-proxy.toml you read the comment and go to a thoughtfully left link that shows you available dns servers right now (https://dnscrypt.info/public-servers/) you type in 'cloudflare' or 'google' or whatever in the config and then restart dnscrypt-proxy
just like that, you have an encrypted connection to whatever dns service you chose to use :)
on windows, by the time you could have already changed your dns, you'd be searching for the right powershell command or finding the goddamn option in the control panel. and the best part, that'll be a regular, insecure as fuck plain dns protocol that goes to 8.8.8.8 or 1.1.1.1. LMAO all this to view catboxes btw.
>>8883993 >still doesn't get it What you want isn't simplicity, you just want the answer regurgitated for you, no thinking required. That's not simplicity, that's someone else doing it for you, just like the worst of the zoomers. Whatever happened to curiosity, discovery, figuring things on your own by fiddling with them until you understand how to work with them? GUIs enable all of those things since they're easy to use and let you see what your options are; CLIs are anything but simple when you actually try to use them earnestly, in fact they're SO overly obtuse and hard to approach that you practically have no choice but to Google/ask an LLM. What happens when you don't have access to those things? And it really goes beyond just GUI vs CLI, it's emblematic of one's approach to things in life as a whole. Also lmaoing at your assumption I use cumrag, I hate using that shit for vidgen with a passion because it's only a step above CLIs and avoid it as much as I can.
>>8884006 I went with a tattoo that attracts horses instead. >>8883996 Why is her breasts coming out of her stomach? >>8884007 I'm not reading all that shit, I already have simplicity, pasting in literal 2 lines into powershell which I can open instantly instead of going through 10 steps. Going through those 10 steps isn't "thinking", it's repeating actions that someone told you to lmao.
>>8884012 >"i can't hear you blah blah" >someone told you to (yet more projection after I spent an entire post explaining how GUIs are navigable without external guides unlike CLIs) Concession accepted.
>>8884007 >CLIs are anything but simple when you actually try to use them earnestly, in fact they're SO overly obtuse and hard to approach that you practically have no choice but to Google/ask an LLM. >>8884014 >GUIs are navigable without external guides unlike CLIs nta but you're wrong on this one. that's true for powershell, but it's not the case for linux at all. once you learn what basic commands (ls, rm, mkdir, cd, cat, echo + your distro package manager and init system) do and learn of some easy text editor like nano, remember that you can use up and down arrows to scroll through past commands and tab to have interactive suggestions, navigating the cli is very easy. almost all linux configs reside in /etc/, you literally can choose to explore anything you want in there.
I haven't been to these threads in a while and I will not scroll through your schizo shitposting, bumbling retards and nigger posting
What do you guys think of the Anima release? Was expecting it to be a little easier to work with than the previous ones but it almost feels like a placebo
>>8884061 >What do you guys think of the Anima release? Was expecting it to be a little easier to work with than the previous ones but it almost feels like a placebo Yeah, prompt adherence didn't improve that much, style is a little more consistent across gens now but you still need to roll a very good seed for it
>>8884066 I don't know what that means and I do not care seethe and dilate faggot >>8884068 >little more consistent across gens now but you still need to roll a very good seed for it Fuck, that was the thing I hated the MOST about Anima, maybe even more than not having inpainting or artist mixing Using noob I can just type in shit and it spits out a good, consistent gen, and the lack of tags for specific things and natural language were the only issue (aside from the atrocious VAE), but Anima can't do that for shit Every single prompt and artstyle is inconsistent as shit
>>8884103 How do we know which one is correct if you didn't even include the artist? Huh, retard? Left is cuter, but that doesn't help if maybe right is what it's supposed to give. Heck, maybe both are wrong lmao.
>>8884116 Yeah, we all expected it to fix many issues with the previous models but it's pretty much just a scaled up preview 3 Seeing as I'm brown, I suppose I should continue using sdxl
>make gen >perfect replication of the artist style, but it has some error in anatomy or composition >hit gen again >good anatomy but the style is completely different
Even after weeks of spam, /edg/ is more lively and willing to post gens than this shitty place. Hilarious how you cucked faggots blamed baker anon for all your troubles and got yourselves the worst AI thread on 4chan
>>8884182 /y/ is unexpectedly full of faggots. Or maybe I should say empty of faggots since no one goes there kek And even if you ignored the type of people that post there, most of them don't know shit about anything and just post shiny retarded slop
>>8883709 The better the model, the less discourse is needed, the less questions are asked, the less anons need to deal with this little brown bro hole. It tracks from all the way back to SD1.5. The majority left are just here to shitpost. It's just in different capacities.
hello everyone, please play this song starting at 3:57 before reading the rest of my post to set the proper tone >https://www.youtube.com/watch?v=Cwcinb2OxUo is there a way to define an END range to a prompt with prompt editing? lets say i want the tag otoko no ko to be only active for the first 75% of the steps and gone for the remainder, is there a way to do that?
>>8884208 What are you on about he's very cute >>8884210 Based >>8884211 Based too >>8884212 [otoko no ko::0.75] >>8884214 Ewww I can do loli and femboys but I'll never combine the two
>>8884229 I like both varieties But particularly dislike it when they have no shoulders and comically large hips Notice the v-taper in that gen Now enough faggotry before I get banned lol
>in an explicitly straight thread ahahahaaha is lil bro serious you in a hentai thread with a bunch of dudes touching your penises together sounds pretty faggy to me
>>8884269 Not me, I'm just pointing out the reason why the faggots push the "it's not gay" angle. They don't really want the company of other faggots, they want actual people heh.
I merely dislike dogs. (You) want to see dogs having sex. Hell you put dog anus front and center in the gen unironically. Out of these two options? Yeah I'll pick being me any day!
>>8884307 Bro...... the horse on horse gens were ironic if you didn't catch it. Also a test of what could be done with noob, it wasn't easy. But the horse on girls gens? Yeah those are nice, but I would never fixate on a horse anus wtf?
>>8884316 You'd know if I was really into it though. I would keep posting it and become another centrifuge tranny except it would be horses instead of faggots!
Got kinda curious what would it look like if I prompted for fat girl with @butterchalk, but not curious enough to type fat in the positive prompt box (never done that in my life, 6-digit amount of gens generated).
holy shit holy shit https://civitai.red/models/2385278/animayume?modelVersionId=2963515 >This version was trained on Anima Base v1.0 using my custom dataset (a mix of a small e621 dataset and Danbooru). In this release, I added many new characters and improved the existing ones. I also enhanced support for various artist styles, allowing the model to generate results that are much closer to the original styles. In addition, the model now understands some concepts and knowledge from e621, although the support is still limited.
>>8884516 Why are you linking realistic gen? Pics with detailed shading used to be the standard here, but for some reason everyone's posting shitty flat-colored sketches now.
>>8884517 it started with nai shills trying to convince everyone that anime screenshots and flat colored crap is top aesthetic because their models can't generate anything else. one of their most successful psyops
>>8884524 I wonder. What is with this? I've been playing around with anima and adding/removing the score tags and shit is wildly inconsistent. With artist style, it has almost no effect. But without specific style, it means difference between barely recognizable scribble and well-looking picture. Or was something wrong with my tests?
>https://civitai.red/models/2505864/pornmaster-anima-preview?modelVersionId=2898873 Without sounding schizo, explain why you're not an anima master of porn?
>>8881901 >be ayymd user >want to use anima >everyone says: use forge neo! >clone the repo and do the Linux install method >doesnt work >retry a few times >doesnt work >ask gemini about if that shit even works for ayymd >"sorry pal but looks like the developer made everything to support nvidia only, i checked the python scripts, also it says right there he doesnt supports any ayymd error" FUCK ME is cumrag ui the only solution for me now, please god no, HELP
>>8884607 set up forge neo python environment manually inside rocm docker image, it should work you can use something like this as a template https://github.com/IgnatBeresnev/comfyui-gfx1151
How would the bright minds of /hgg/, the AI general for the thinking man, replicate this kind of style on Anima? 3D and Photorealistic are not strong enough to create the kind of soft lighting and skin detailing it features
>>8884633 this is only possible with tools like novel ai, if i was to try this with anima i would wait a couple years for the model to go through a few finetunes
imo anima is good but definitely needs some kind of aesthetic tuning, sad to see that 6 fingers or more is still a thing and qwen is dumber than i thought
>>8884639 I feel like it would require a massive amount of editing/post processing And you still need a base that features the kind of realistic anatomy and realistic textures I think I see some quasarcake in it, but it's too painterly so it must be mixed with something else Unless the guy first gens it with a more realistic model and inpaints the head
>>8884639 >>8884641 I guess the actual question is, could you train a lora on a half dozen of these and make a stable style out of it. Nobody wants to do this amount of post-processing by hand.
why does anybody has compute nowadays guys on lax's cord server been saying anima is a thing only because comfy was on their side and trained the model literally there's no other relevant model rn
>>8884651 you could definitely make a lora from the guy's pictures on pixiv, he's got 204 uploads and multiple videos to grab shots from but personally the fun is figuring out how to create something similar myself I just love artist mixing and the creativity that comes with that to me it almost looks like he uses a 3D rig and inpaints over that as I just don't see how you could squeeze this kind of detail out of pure AI gen
>>8884652 >anima is a thing only because comfy was on their side and trained the model Is this supposed to be some secret or revelation? It's been advertised as the first model bankrolled by Comfy the org since day 1 with more models to come in the future. I remember being happy that the local community finally found a way to fund new models in a centralized, sustainable way and that we weren't going to be relying on random chinks of furfags making one-off models with the one chance they get through any variety of means, then disappearing forever. Did this announcement fly over people's heads or what?
>>8884656 true though ive heard theyre waiting for some miracle arch/model so lax can throw a large amount of compute to it z-image was not it, klein was a maybe but no, sensenova is too big, hidream was impressive but i dont think theyre convinced
This is a backstory and "first scene" from my r-18 story I'm messing around with.
>When exploring the castle during night, looking for secrets. [Princess] Aurelia, accidentally stumbles upon a party organized by a corrupt [Nobles]. There, a drunk [Official] mistakes her for a prostitute and rapes her. Afterwards, she manages to stumble back to her room, and no one else would remember. But this experience awakens something in her. And a irresistible craving forces her to learn more.
>>8884781 You could mask out the problem areas, let the model figure out feet on its own. Or use an earlier end step. CN strength doesn't seem to do much at all with these Lllite ones, there's more of a difference between 0 and 0.01 than there is between 0.01 and 1.0
>>8884878 in case you've forgotten, the existence of /r/ is useful as a catch-all to tell people who want to freeload to fuck off. also the sticky makes specific mention of it but now it doesn't exist anymore.
personally I think anima is fine my only problem is that it has learnt artist styles TOO well, to the point where it insists on using features from old styles artists have abandoned and you can't neg out something the artist does that you don't like
>>8882633 I baked with a LR of 0.00002 and it seems ok, however it has body issues where it'll make the image too thick, long torso, or both, do I raise the LR or increase the steps? it's currently at 3200 with 40 images. (batch 1, 80 epoch)
>>8884995 That's another anon. There was "original" miqo, the one who stole the design (that's the thread famous one) and than the third one who stole the stolen design for his game.
>>8884929 Pony's release was the actual "explosion", mostly because there was no Civitai yet and we had the biggest lora repository so a lot of normie tourists would stop by and some would stay. Activity has been slowly declining ever since, with temporary spikes for illu and noob release.
The main issue is, everyone who wanted to gen things has been able to do so for like three years now. Being able to make free art of your girl/fetish is not as exciting anymore.
>>8885032 Civitai was there since the 1.5 days, anon. Pony was a huge explosion not because of loras, but because it was the first time prompting for "desi big bobob ladie" produced acceptable results. Then people just came flooding in because they got banned from discord for posting Zelda or something.
>>8884929 Pony was unironically the biggest jump to date, it's hard to understate how much better properly trained XL was compared to 1.5. Illu/Noob were still XL so the core functionality remained the same as Pony, just with better aesthetics, but said aesthetics were the best ones up to that point by a long shot, and arguably to date. Anima, comparatively, didn't bring THAT much new - NL is nice and I really appreciate being able to do more complex compositions, but in a world that's familiar with Nano Banana and GPT-Image, it just wasn't very impressive even when the first preview dropped, let alone now at release.
>>8885033 Hardly the first time, the same would have worked on Animagine in SDXL resolutions. And porn poses already worked on EasyFluff, although that only got like two months to shine before Pony release.
>>8885038 anon he's trying to shillpost about nai v3 >>8885039 >animagine censored dataset and it came out AFTER pdv6? Or at least in really close proximity, something like within 3 days. Also it was really fucking bad. >ef lol lmao and you're probably going to pretend hll somehow aesthetically fixed that piece of garbage.
>>8885039 >Animagine NSFW-incapable, even at SFW it was shit. >EasyFluff Basically unusably bad. XL really wasn't much of a thing until Pony for a reason.
>>8885113 >umm ackshually that's against our content policies >doesn't direct you to .red fucking cuck website. at least the childrens have been saved, thank you civitai
>>8885120 I haven't frequented the threads in a long time and I didn't know. When I googled it the results didn't include the new domain and instead showed reddit threads, but those require you to make an account to evensee the replies.
>>8885127 nta but I don't have that. Think it goes away the first time you dismiss it, like other announcements. ...which would still make it anon's fault
>>8885132 I am NOT going back to a model that struggles to generate the most basic of FFM positions and gives the guy a strap-on when prompted for one on a female.
>>8882633 it works great in a hybrid normal prompt + sentences esque of prompt but with the overall image keep being dark, (blue/gray dominant), the bodies always have problems with longer torso and/or tendency of thicker body, does anyone know which to remove from the setting?
>>8885144 I have no idea how much to put or what else to change in the settings, the images are also sometimes are cropped off, i've set bucketting to: 256 min, 3072 max, 64 steps. I baked with the base 1:1 recipie of that toml and it still suffers that fat/thickness/elongated of a body minus the head. The loras also have half the filesize that i'm used to, this one is ~89mb other I have are ~175mb to 250mb
>>8885154 you shouldn't use batch size 1. set bs to 2 or 3 or 4. use gradient accumulation if you can't use real batch size. lr 2e-5 with dim = alpha is fine. i use lr 4e-5, alpha = half dim, batch size 2, for 40 images i would train for 30-50 steps per image. if the lora is overtrained, use network dropout = 0.001 to 0.01, reduce the training steps, or use earlier epochs or you can use adamw8bit + cosine/cosine with restarts, lr = 1e-4 which is a safer choice than came rex
>>8885157 thanks i'll try that out, i'll raise the batch size and set network dropout to 0.001, does 32 dim 16 alpha on both go better though than said dim=alpha?
>>8885159 >32 dim 16 alpha on both go better though than said dim=alpha doesn't matter much, if you change the alpha/dim ratio then you should scale the lr. dim=alpha is fine
>>8885161 i've set it to batch 4, 2000 steps and the timer went up to 2 hours compared to batch 1 3200 steps. For the sample args where it'll print out an image every certain epoch/steps, is the txt file just the same as txt tags in the tagging or something complex?
>>8885167 > is the txt file just the same as txt tags in the tagging or something complex i have never used this, but i guess it just a normal prompt file for sample images > batch 4, 2000 steps yeah sorry, when i said "for 40 images i would train for 30-50 steps per image", i meant 30-50 repeats per image, for example 50 epochs 1 repeat or 25 epochs 2 repeats
>>8884633 I think I've gotten as close as I'm going to get without doing a shitload of post processing in photoshop first one is with no loras, just my artist mix and comfy setup middle is https://civitai.red/models/2628200/anima-lighting-atmosphere-enhance?modelVersionId=2950803 third is https://civitai.red/models/2459979/light-concepts?modelVersionId=2952873 it's very hard trying to balance the 3D style for the body, more natural style for fabric and the flat 2D face, and I don't think it's super consistent
>>8884802 yeah there's a certain appeal to mega filmic super slop, it just works as it goes from being overdone to a neat style
>>8885171 it's an improvement, stopping at 700 steps, but just like the other rawr/came toml, I need to raise the weight to as high as 1.5x to get the right results, is it intended to be this way or am I suppose to raise the LR?
>>8885188 does batch size degrade positions or is it fine to raise up? I feel like I can set it to batch 6, otherwise maybe just batch 2? Also the images still do look very dark, not sure if its the 0.99, 0.999 settings there, otherwise it's getting better every new bake.
>>8885234 >>8885235 >I did notice the "full bf16" is unchecked full bf16 and mixed bf16 training are both fine >degrade positions are you training a concept lora? >batch 6 batch size 6 is also fine i guess, i just like batch size 2 results better. just make sure to scale the lr when you change batch size and alpha/dim ratio https://github.com/spacepxl/demystifying-sd-finetuning
I keep getting weird artifacts trying to run novaAnimeXL_ilV190 (Illustrous), I don't usually have any problems like this so I dunno what to do, anima still works fine
>>8885248 >batch 2 3e-5 with alpha = dim, or 4e-5 with alpha = half dim also i use the default betas for came, which is 0.9, 0.999, 0.9999 and weight_decay = 0.08
>>8885279 >Have you ever been to a dedicated NTR thread No, and I have those all filtered out I'm obviously talking about non-dedicated threads >>8885280 No one would do that if they didn't like it already
>>8885301 Link? >>8885302 Doesn't work. It never looks like I actually inpainted for some reason. Maybe I'm using the wrong settings, would you mind sharing yours? Sorry for the spoonfeeding
>supposedly the most high skill AI general of 4chan >can't even inpaint with a model that uses the exact same inpaint workflow as every single other local model of the past 3 years
>>8885318 can you not be a nigger and post your metadata so we can tell what's wrong with your settings? Besides, it seems to have inpainted just fine, just at a low denoise.
>>8885341 I copied the settings from that anon, exactly as they were. So the denoise in here was 0.5, which is what I used on SDXL too. This is another one but here's the inpaint: https://files.catbox.moe/bb3tpx.png
>>8885349 while i haven't tried using the inpaint cn for detailing, i found it to be very picky about it's sampler settings. don't use anything besides er sde when inpainting
>>8885350 In this thread it's prohibited to make use of things when they fit >>8885352 I messed up with the settings somehow so I disabled it for this one lol >>8885353 Adjusted to my hardware faggot >>8885355 Hmm I see, I'll change the sampler then
>>8885358 No, but I didn't really think those would be massively consequential Higher shift looks ugly as fuck though, and er sde tends to go for a more flat type of shading, which isn't really what I wanted for this one in particular But I'll give both settings a try in a min >>8885360 Not particularly, depends a lot on the compatibility between artstyles Some are mutually destructive in my experience >>8885365 >assuming it's the sampler already 4chan indians are presumptuous like you wouldn't believe
>hi guize how do I solve this issue? >you did X wrong >wtf why are you assuming X is the cause of my problems you jeet Why even ask for help in the first place
>>8885368 >>8885374 Yeah but turns out er sde doesn't work either so it's obviously not that you precocious clit God imagine being so retarded, so wrong and so conceited at the same time Makes me want to thank all the anons that weren't autistic
>>8885376 I can't account for the jeetvitai loras you are shitting your gen up with, and for you reducing the shift to the point that the model can't function properly
>>8885379 That's a screenshot retard, furthermore the face is mangled BECAUSE of the inpainting The base one is the left in this >>8885376 Notice how you stopped pretending the sampler was the issue and changed the goalpost, because you're a presumptuous indian just like I said Now I will wait for someone of good will and preferably not brown to help me with this
>>8885383 You aren't even prompting for anything specific there, what are you expecting to change in your post stamp sized close-up? Whatever possible variance in output is probably being killed by your jeetvitai lora.
>why do you guys still try to help this terminally retarded poster? because i am super man https://www.youtube.com/watch?v=rpO1U-nEgRU >>8885349 for detailing use the anytest cn instead, inpaint cn is for adding/fixing things at a high denoise
I was getting Claude to teach me how to use Stable Diffusion and it was happily going along with it, giving me prompts to get what I want helping me build this NTR image, suggesting its own ideas for turning it into a set. I ran out of tokens and asked it to make a prompt to carry over the progress. When I began another chat now Claude refuses to do it anymore, saying it won't help make prompts for adult images, especially ones with dodgy consent (claims NTR is no bueno). What gives? How did I even skirt around it in the first place?
>>8885411 Beats me, but it honestly did a pretty good job. It was onboard with the whole sleeping wife cheating on her husband, made all the postive/negative prompts for it. Told me exactly how to configure everything, walked me through inpainting the husband into the background, came up with a morning after scenario in the kitchen on its own, showed me how to reference the original with ControlNet to keep the same design. Basically said it was down for anything except loli. Now it won't help me at all with anything adult related. Was using Opus 4.7
>>8885248 this one is an improvement but not sure why its twice as slow. the toml has ZSNR turned on and I also ticked that off for this one, could that be why?
>>8885456 Because you're doing more steps. Batch lets you utilize more vram by processing more images at the same time, it won't cause a slowdown as long as you do have the spare vram.
Also your optimizer config changed, betas. No idea what that does t bh but I wouldn't expect an impact on speed.
weird how the AI threads usually die at this hour 1am for euros, so perhaps neets going to sleep? 8pm east coast US, so burgers stop phoneposting from overtime work? china and japan should just be waking up, india still asleep,...
>>8885463 >>8885462 Thanks guys and that guy I pestered since morning. Considering it takes me 2 hours to bake the last one so i'll do a bunch overnight later with other sets, the 0.99,0.999,0.9999 thing is based on the other guys suggestion, but next time i'll revert it.
https://civitai.red/changelog?id=246 >This week, Members get early access to Anima training. If you've been waiting to fine-tune in this ecosystem, you can now. >We're retiring the Kohya trainer. We've fully transitioned to AI Toolkit so jeetai added anima to ai-toolkit and didn't make a PR to upstream. cool
>>8885671 >only one of those four draws greasy skin The greasy skin is mostly because of the various greasy skin tags like "shiny skin" and "very sweaty" >Most aren't even in color Out of those, only otimusyairoha mainly draws b&w, and Noob could alreadt draw colored pictures in the style of monochrome artists (with some funky shading)
And you want me to post my cute polished gens, each of which is a literal masterpiece that puts Mona Lisa to shame, in the same thread with these subhumans???? No thanks, I'm not posting shit until the internet is segregated again!
>try to replicate Anima example images from Civit >looks completely different in cumrag >looks completely different in another way in Forge Neo What the fuck? I remember doing the sanity test with the one preview1 example image from HuggingFace back before any of the WebUI derivatives supported Anima and it came out as expected.
There are some things outside the UI that can prevent full replication, fast fp16 accumulation comes to mind. I also remember reading about how the way the seed is translated into noise can be affected by hardware.
>>8885743 Thanks for taking your time, but I went out my way to disable all the optimizations by removing them from the command line arguments to no avail. RNG source really is the last thing I can think of, but I've no idea where it's set in cumrag. AFAIK CPU RNG is consistent across machines whereas GPU RNG is hardware-dependednt, and what I fear is that CPRussel used GPU RNG, in which case it won't be possible to replicate the examples at all unless you just HAPPEN to have the same GPU as him.
>>8885746 That would also be terrible for anyone trying to pass the sanity test, and frankly pretty unprofessional too. I just want to make sure I'm getting the images I'm supposed to be getting before I layer on optimization after optimization.
I didn't think people would get so pissed off at three whole catboxed gens, but I guess that's more pictures than anybody has posted here in the past 72 hours.
>>8885782 I imagine it's not that easy to do, not to mention do it safely. People have been looking for dick growth medicines for thousands of years and the best we have is still just a saline injection.
>>8882633 I baked multiple things with this and all of them came again, darkish, grey-blue bias colors, is it just the restarts causing it I can't figure out which one, I don't think its the vae unless theirs a problem with eztomls, and I tried other method such as glora type of baking and the colors come good.
>>8885784 Allegedly, in the Ancient world small penises were a sign of upstanding moral virtue and intelligence, so we lost a headstart of thousands years on penis enlargement medicine.
>>8885795 >Allegedly, in the Ancient world small penises were a sign of upstanding moral virtue and intelligence this is because when you peel back the foreskin of a small penis you find the cheese of the gods. a supreme delicacy that improves with age
>>8885806 Interesting. The whole issue of mixing artists within Anima is turning into extremely complex issue. This is third way of doing things. And I still don't have working way to quickly test combinations.
>>8885806 why would he not provide any examples of its effects? also it doesn't sound like it's compatible with prompt control https://github.com/asagi4/comfyui-prompt-control and I think I'd rather have scheduling than testing some randoms artist mixing
>>8885806 this is what his style mix would look like with his prompt and parameters but without his funky artist mix node, because I'm not installing that shit I guess it actually does something, but I'm gonna wait for someone else to vet that shit first
>>8885853 Nope, I just got in here. >>8885854 Of course BREAK doesn't work, that was an example. It's just that it seems like this is more or less what this extension is doing.
>>8885863 In the time you have written your opening statements, Claude has already committed a whole update and deployed a test environment to sandbox the changes. You are going the way of the morse code operator, I hope you had fun wasting your life training for nothing!
>>8885866 >Google says 75% of the company's new code is AI-generated https://www.businessinsider.com/google-ai-generated-code-75-gemini-agents-software-2026-4 >Microsoft CEO Satya Nadella revealed that between 20% and 30% of the code inside the company's repositories is written by AI tools like GitHub Copilot https://techcrunch.com/2025/04/29/microsoft-ceo-says-up-to-30-of-the-companys-code-was-written-by-ai/ >Oracle says AI coding tools are helping it dodge the SaaSpocalypse https://www.theregister.com/off-prem/2026/03/11/oracle-says-ai-coding-is-helping-it-dodge-saaspocalypse/5225909 >100% of NVIDIA’s software engineers utilize AI-assisted coding tools. While measuring the exact line-count of AI-written code is impossible, internal metrics show that over 30,000 engineers use these agents (like Claude Code, Codex, and Cursor) to check in three times more code than they did previously. https://www.linkedin.com/posts/nvidia-ai_100-of-nvidia-engineers-code-with-aiand-activity-7414335445246902273-ldvG >claude code will probably make 50,000 people millionaires if not more https://x.com/gregisenberg/status/2007829847796863409
Why do you refuse to move to the next step of human evolution and stay poor? Are you stupid?
I might do some actual XYs later, but it doesn't feel any better than scheduling so far. Also the: >Style-similar artists mix well; style-divergent artists may "regress to the mean" into a compromise look. note is not encouraging.
>>8885785 nice >>8885891 why do cumrag users on average have uglier image posts than people using reforge? you'd think all the noodles would actually be used for more interesting workflows to fix limitations with reforge's simplicity
>>8885901 A cumrag user is either a retard who opened the first link he found on google while searching for how to gen hentai and got sent to some Indian's tech blog shilling his paywalled workflow, or a /g/ tranny who only cares about having the most cutting edge nodes and lacks even the bare minimum of aesthetic sensitivity you need to gen something nice to look at. Add in the fact that 90% of people in these threads (and /edg/ in particular) never cared about anime fanart in the first place so they know 5 artists tops and have no idea of how to visualize the kind of gen they want to make, let alone come up with a style mix.
>>8885901 Honestly I just pulled up a gen that was ready for a comparison, don't have time to put in effort now. The original image looked better than with this mix thing, but that's sample size of 1.
>>8885906 >Add in the fact that 90% of people in these threads (and /edg/ in particular) never cared about anime fanart in the first place more anime haters to hate, perfect
>2026 and cumrag users still can't into xy grids kek honestly if you're not doing vidgen or trying to automate something that's novel using cumrag makes you such a retarded tool
What are you anons talking about? I'm clearly the best genner around here and it's not even close, and I use comfy. Oh right, I forgot, I haven't been posting any gens for centuries!
>>8885909 Efficiency Nodes did more total damage to Comfy users than when ModuleManager was distributing malware. When doing XY all the noodles are fake and handled by a script behind the scenes. The noodles that would have worked with any model are locked to SD, and also incompatible with diffusion-CG, IP-Adapter, Controlnet,or any other tools. And there's a bug with lora XYs, where after the first lora it sets the TE effect to zero for all the later ones. It's been there for as long as I remember, probably before Pony, and nobody ever fixed it. Most of its lora XYs are straight up wrong.
That said, the alternatives are enough of a pain to use that I just gen the variants manually and use a GIMP script to setup the chart.
>>8885806 Tried a bunch of mixes that looked alright on noob, with artists that aren't too similar so it's easier to see what comes from where. Left to right: @artist1 @artist2 @artist1, @artist2 [@artist1:@artist2] anima mixer
I legit cannot see any pattern in how they mix. Probably going back to scheduling where I have more control, even though it makes the mix unstable between seeds.
After playing with it for a bit, it definitely feels like the AI is just inside some black box it uses to recognize artstyle and combine things in way that aren't obvious to outside observer. Now that I'm thinking about it, I did experience things like that when trying to combine authors.
>[[artist1|artist2]::0.XX] [:artist3:0.XX] (3 or 4 artists is the sweet spot) >negpip -> old, early, mid (if you value consistency) >RDBT at really low weight (no composition issues ever) >only Best Quality OR score_7(HELLO SIR) in POS >ER SDE/Simple for harsh styles, Euler/Normal for soft styles, CFG 3.5 It took a bit of fiddling but I finally wrangled this little nigga perfectly. Shout out to anon for negpip and scheduling. Glad we LET TOUCHDOWN RUSSELL COOK.
ban all phoneposters if you use a mobile viewport your IP should immediately receive a ban smartphones were a mistake steve jobs is being tortured in hell over this, among other things
if every 5 epoch saved constantly changes things such as hand position, expresstion, any type of furniture/item position, etc. does it mean it's still underbaked or is it because of the "warm restarts" type of training?
>>8886105 nope, i'm trying to bake the best I can. constantly changing little things and the pc has been baking nonstop. I would try something that's better than that toml combo though.
>>8886116 >this is the original setting before I swap to lycoris well then you've already used locon. it's not lycoris's locon, but it should be the same
>>8886156 Explain how>>8886060 is Western in style (?) or origin (every single gen is of Western origin because they are based on American software and Hardware)
>>8886170 Why wouldn't you use negs? How else are you going to tell the model to avoid stuff like bad anatomy or chromatic aberration or niggers? Genuine question.
>>8886171 >Why wouldn't you use negs? Because they don't work the way you think, it's not a magic box that tells your model to not put something in the picture. Learn how the tech works, then come back to have a proper discussion, I'm not your teacher and I have no intention to entertain your ignorant ass any longer. But you should stop spreading misinformation based on your lack of knowledge of the math behind image gen and make people commit the same pedestrian mistakes as you.
>>8886203 img2img artifacts from some anima architectural quirk either raw gen into high res to avoid upscaling, or use tiled upscale with tile size below 2mpix
weirdly enough i would expect multidiffusion to also count as doing tiles, but clearly not
>>8886203 Post a catbox or it's impossible to tell what's wrong with your settings. >>8886205 >>8886206 >>8886207 I use mixture of diffusers and never had those artifacts, the problem is somewhere else.
>>8886207 Actually, said pic was generated with high res fix x1.4 then upscaled x2 with mixture of diff, the tile size is always at or below 2 mp in my settings though...
>>8886213 It looks like switching to 4x-animeSharp removes most of the artifacts. I've been using 0.1 denoise because I feel like a higher value kinda messes up some of the details (especially the eyes)
>>8886216 >4x-animeSharp It's REALLY old and pretty bad now, I'd recommend moving over to 2xPublic_realplksr_dysample_layernorm_real. It destroyed 4x-animeSharp in a back-to-back test I did a few weeks ago.
Tired of all the otoko gaslighting from non otoko genners. Where have all the miquella gens gone? The smtv mc? The inazuma traps (formerly otoko_no_ko)?
>>8885655 >[@ningen gokko, @deadnoodles: @shiroi yoru, @otimusyairoha:0.333] What's this syntax, in particular the colon after deadnoodles, but also the square brackets. artists changing 1/3rd of the way through generation?
>>8886292 nta but yes. that syntax has >@ningen gokko, @deadnoodles for the first 33% and >@shiroi yoru, @otimusyairoha for the remaining part of the gen.
>>8886294 I wonder how people come up with these combinations. Do you just randomly throw stuff at the wall and see what sticks? From my experiments, author combinations are more random than any artistic melding.
Especially when there are multiple artists, in different orders, in different ratios.
>>8886296 you test it? but even beyond that there is logic behind it. if you want the foundations of <a> but don't want it to for the full way through then you can [<a>: :0.2] or if you want the aesthetic of <b> but if it's actually in the foundation of the image it comes off too strong you can do [ :<b>:0.3] in this case it's wanting the anatomy/compositions of X artist(s) but the actual linework/detail of Y artist(s).
it sounds a lot more complicated than it actually is.
>>8886296 >Do you just randomly throw stuff at the wall and see what sticks? p much. However, like anon >>8886300 pointed out, there is a method to the madness. Even chatgpt can break it down for you but the first artist (or two if you alternate with each step with |) serve as the foundation and composition while another artist in say the final 25% of the generation can work to straighten all up. Works great if there's an artist with a soft style that can produce fucked limbs but you can add stability to it with an artist with a well defined outline. Also say there's an artist who predominantly draws fuck huge hags but you want to pull their facial style, you can prompt an artist primarily known for petites for the first few steps then switch out to the former artist. It gets the foundation and then just fills it in. Ideally, you'll never have problems with 3 artists this way,
>>8886332 im not into pussies so i do not want to be shrunken down and sealed (using bar censors as jail bars) for all eternity in the cavernous depths of her fishy vagina
>>8886332 her anus on the other hand... i need to be shrunk down and sealed into her rectum, where i will live the rest of my sorry brown life as a janitor for her intestines
>>8886349 my only guess is they're relying entirely on the character transfer to insert the girl and are trying to skew the uh, composition from the shitty reference.
>>8886399 >There should be a line beneath the "01," but it's missing. Dayum not every artist adds the "line" under the number. I thought the complaint was the extra dot in middle bottom of number 0, which *is* unconventional.
>slop the sloppiest nais slop ever seen by humans and machines >open clip studio paint (ms paint) >edit a pixel >upload to xitter LOOK GUYS IT IS 100% ART I ONLY USED AI FOR THE BACKGROUND!!! >:^(
>>8886417 Care to post your pixiv and twitter account so we can compare it to Ayakon's portfolio? You must be a popular AI content creator to throw shade like that, right?
>>8886436 Yeah, like, he gens pathetically bad slop and that makes his attempts to sell himself as some kind of brand laughable. We get it, we got it months ago and he hasn't posted here in ages. No need to try and make hims some epic lolcow.
>>8886445 >build a following for commissions >share consistent style portfolios >get instant feedback & trends >network with other AI artists >showcase high-res outputs
>>8886453 anon are you afraid that you won't be able to get any attention if you were to go to pixiv or xitter? Do you have no faith in anything you do? Are you afraid of being worse than ayakon?
what we need is some sort of collection of the best gens of each thread to make genners feel like their effort is being recognized and encourage people to post high quality gens instead of slop.
>>8886475 but we should also include lesser quality, but clearly made by new to the thread posters gens to stoke an ego boost and increase their activity levels
>>8886476 You are right, and to avoid petty jealousy we could have the gens not ranked or arranged in any particular order and instead have a nice collage that puts them all together.
>>8886478 let's also make sure that they're downscaled in a very mediocre and bad fashion that actually lessens their quality, this allows the not so good gens to stand out less and encourages people to go back to the previous thread and find the source of them
A good bake of a lora in Anima is ~2000 total calculated steps right? Or could it be more/less? In noob 2464 is the true end point to me. Any higher does nothing. Assuming picture set of 25-40.
>>8886502 thanks, came is odd, it does unique positions and variation decent but it somehow can't recognize brush and linestrokes well as it's going to the point of generic esque of highlighting, while adamw is very true to color and linework as it's very baked in but suffers same position or missing an arm or so, hopefully someone figures out something new or more in the in between the two, that or i'm becoming more obsessed with perfection baking.
>>8886491 Anima is a bit easier to train so you can dial it back a bit. My Style Loras are about 1600 steps while my character loras are just 400 steps, but I use Came/Rex for my optimizer/scheduler.
>>8886504 >hopefully someone figures out something new or more in the in between the two Those are just optimizers bwo, maybe try to to improve your dataset first
>>8886506 yeah, i'm using that toml from above still (those same metadata graphs). Now i'm figuring out if higher or lower LR helps. Issue now is the face is good, but the body isin't, at higher end of the steps the face is "samey" while the body is near accurate to the style, including good hands. Lower steps have better face expression but worse body/hands so I have to add some padding of sort? Adamw is again better at the body/face blend but I do unusual testing like isometric, top-down, worm's eye view/low angle, adamw will suffer at making it unlike came. I seriously hope regional prompting comes, that's so good in noob, and might be better in anima.
>>8886509 I honestly might, I gotta redraw my psd's, perhaps the aspect ratio is the issue, despite being "almost" good.
>>8886521 tried it a bunch, it'll generate something genetric like 1girl, solo, I can do NL prompting of positions but it's inaccurate and has limits. Unless someone has a workflow for it, it's odd but that no head black/white girl with the metadata catbox in the previous thread ultimately fixed my problem with inpainting, maybe i'll wait for that guy again.
>>8886525 Do your NL prompting like normal, apply that prompt to all regions. Then start pruning the stuff that leaks, from the regions where it's not supposed to be.
when ai art is finally accepted as real art, i'm going to pretend like i was fighting the good fight in favor of it the whole time instead of doing nothing but jerking off to my own gens #stolenvalor
>>8886477 X or pixiv for money? Nope, that's retarded. Never monetized shit. I do it so people know how amazingly artistic and creative and technically proficient pedophile I am.
>>8886603 Unironic answer: yes. It's slower (unless you use turbosloppa LoRAs) and a bit less aesthetic than Noob but the NL capabilities make it easily the best NSFW-capable imagen model right now by a long shot, shit that wasn't possible in a gorillion years on SDXL derivatives and took a hundred tries on NAI v4.5 takes maybe a dozen tries on Anima. It's not magic by any means, in fact it feels pretty fucking outdated once you've had a taste of actual SOTA (Nano Banana, GPT-Image) as did most people even remotely interested in Imagen, but that's par the course for NSFW-capable models.
>>8886541 AI is indeed good, AI artists do not exist though, just like those "artists" that rely on paint tools to "draw", if blud can't draw shit on paper then they aren't anything but deluded.
>>8886615 Ishikei, Koku, Kobuichi, Happoubi Jin, Shida Kazuhiro, Mitsumi Misato, Pochi-Goya - not one old guard artist that's been drawing for longer than 10 years looks good on Anima, let alone better than they did on Noob. Ok, maybe Tsunako, but her style's always been very consistent so she's kind of an outlier.
>>8886617 Just because a bunch of old ass artists don't work well with the model it doesn't mean that it has a worse aesthetic than noob. Noob's base underlying style was beyond dogshit and it seeped through every artist tag without using loras. Anima;s is at least passable and way less biased towards extremely burnt oversaturated reds.
What anons in here mean when they say the base style seeped through every artist tag isn't that the model has an inherent style bias, it's just that sometimes noses are more than a dot
>add flat color to negs >add deviantart to negs >add every score tag to negs >add @incase and @fellatrix to negs >suddenly every gen is immediately fantastic
>>8886637 Ollama is the UI+backend, not the model. You can get the abiterated version of an LLM model, like https://huggingface.co/mlabonne/gemma-3-27b-it-abliterated, which claims to uncensor it with no noticeable quality loss.
>>8886642 I picked the first one that popped up on google, just to make sure I spelled obliterated right. You can pick the model you want, I'm not even sure if using the larger one is better than using the exact 0.6B that Anima uses.
As a result of >>8885515 I will be leaving this place within two days. Wanna say thanks to any oldfags still around. It's been a good run, despite the last few months. I'll miss you retards like you wouldn't believe.
I tried so hard to like Anima but I can't. The lack of e621 puts me off... I can't gen hyper cocks penetrating my anime girls properly. I'm about to give up and go back to Noob. It wasn't supposed to end like this......
>>8886640 I tried running Gemma 4 with KobboldCPP, and them use a frontend (locally-uncensored). But the thing just hallucinated completely unrelated nonsense. I assume you ran it from console directly? Or did you use something else?
>>8886764 I just realized that the average SAAStard has no inpainting and has to either reroll the whole thing or hope the edit model doesn't randomly change anything else in the picture if he wants to fix an error or change a detail.
>>8886802 shin megami testicle. that is what i am. a powerful demon cursed me for my insolence. i got too greedy, begging it for macca over and over again and in return my sorry brown flesh was warped and twisted, turned into random fella nuts. how long has it been? does he even realize, that i am part of him? i cannot see, i cannot hear, smell or speak. only make cums. my screams are SPLURT
>>8886856 I'm just using euler_a/normal like always, but everyone seems to have their own official is er_sde/simple dpm++/Karras is messy as always, if you need more texture details
>>8886856 er sde/simple - Harsher styles euler/normal - Softer styles CFG @ 3 - 3.5 Grab the RDBT p3 v0.24f cfg b weight at 0.2 - 0.3 >>8886860 >rescale Link if Neo? I like a good snake oil or two.
>>8886882 I value a good shitmix shitpost occasionally but this one is from 102d custom nigga. I'll give it a shot out of curiosity as I have 0 issues with Anima Base 1.
>>8886889 The 102d custom guy went ahead to create 102d final which was slopped and a couple others. He clearly has no idea why that one version was popular
>>8886890 >it's meant to share insight into LoRA training & model merging, and to encourage you to merge and post your own. Oh. Never mind then. >>8886892 >102d final Yeah I remember that fumble. >>8886894 >the glossy AI sheen Probably loading up on every quality POS under the sun.
>>8886896 nta but you're not wrong. Then again, are you some civjeet that doesn't use artist(s) and just asks the model for penis in vagene? I get that's why shit like WAI is popular there.
>>8886899 Of course I use artist tags, but whether a model suffers from "glossy AI sheen" depends on its default/neutral style with the officially recommended pos/neg. That's the inherent style everything you do prompt is layered on top of.
>>8886332 Same fag here, I realized that my rig's too poor to train, so I ask if any baker out there willing to train hamada yoshikazu style? I have 133 colored images some are decen edits and already tagged in booru style with trigger word.
>>8886943 >I can suck up images and do whatever I want with them because the fact that the artist posted on the internet = anyone can do whatever they want. If you're protective of your work and don't want people to suck up images, then don't upload it to the net, duh. Many are saying this.
I mean, I would understand the artist's argument if only:
- Artists didn't spend most of their time arguing and thought policing each other. - Good art, dead/inactive accounts. - Suddenly deleting all media from the internet. Or the artist downright dying and their accounts taken down (You are sorely missed, Kogeya). - No archival of any kind. Once it's gone, it's gone forever.
>>8886995 This is true. People are not willing to take more than a few minutes to fill out a form, let alone write a detailed description of what they want. You do not give a client extra loops to jump through. Specially not when some random jackass in the internet can do it for you in seconds.
A reason why most artist's are terrible at business, I suppose.
Hey sir!! Very sadly, you will find that most anons despise bibici coded gens and your image appears to be bibici coded. However, if you redraw that image using your own bibici- I don't care how badly drawn it is, as long as it is still bibici- I can accept it. Hope that helps!!ヽ(^o^)丿
>>8887030 the reason he's thrusing his bibici against her like that is because there is something stuck in his urethra. what is that something? it's me, who has been forcefully shrunk and inserted down into this bibici as punishment. i hope they eventually manage to get me out, otherwise i'd have to spend the rest of my sorry brown life as a pee and cum blocker
>>8886534 this has a low chance of working if at all, unless i'm doing something wrong, I tried making the mask pure green/red or just the normal mask type, noob had it working where you can even add two types, or more of artists of different characters in different artist styles and it won't bleed together at all. hopefully a CN comes or a newer workflow for anima but once this is figured out it's hard to move back to noob outside of better body types and that cozy feeling.
>I tried making the mask pure green/red or just the normal mask type, noob had it working where you can even add two types You can use all the usual Comfy options of creating masks, as long as you don't leave any pixels without a prompt. In fact this whole thing is just plain old regional prompter that would work on SDXL or Flux with no changes.
The point of that post was just my finding; that with NL advanced enough to position characters you should treat the regions more like an exclusion to prevent leakage, and keep most of the prompt global.
>>8887053 it has a high failure rate despite only two regions, the artist even mix together. On noob I literally could make 6 regional prompts with ~90% chance of working, or complete function if only two. Also I don't usually upload images I don't know the catbox upload or setting it up even if I have the script for metadata.
>>8887066 Unironically never, just like artist mixing, shitmixing was only possible due to CLIP being so rudimentary. With an LLM TE, any shift in weight alignment without appropriate captioning will deteriorate its NL capabilities; with CLIP, there wasn't much to deteriorate in the first place. Like, it did happen, but the alignment was so vague in the first place due to CLIP being a glorified word salad that moving it a bit wasn't really perceptible.
>anima was supposed to finally put local on parity with NAI >instead it fucking killed any kind of future development because they chose a dogshit technology to train on kekypow
>>8887076 The upsides of an LLM TE, as well as its potential for growth in future models, far, FAR outweigh artist mixing and shitmixing. Consider those things happy accidents and move on from them, they're never making a comeback.
>>8887081 Thanks for making me discover that my raped goyslave government banned gelbooru. >>8887079 He's right but I don;t really give a shit. Besides, all the "validation" I get is some guy ITT replying that he wants to be shrunk and live inside some girl's asshole.
>>8887087 Tags are excellent, I much prefer using them, but they just aren't sufficient for describing complex compositions. I can only hope they don't get phased out eventually by out of touch people with compute, but NL certainly isn't going anywhere either.
>>8887053 >>8887060 I give up on pursuing regional for now, theirs a limit with how much we can do with the NL/tag mix but it can also do things that's better than noob, it's better to do the waiting game on regional at least for now.
example is just normal prompt, no regional or inpaint/edit. the regional never really worked well for me, but once we do get one though fuck yeah.
>>8887115 Thank you, have a rare image of a tea girl without implants. i hope can improve this style on anima once i make the switch >>8887130 stealth png should work, but here you go https://files.catbox.moe/v9pjtq.png
this one? i'll give it a try later, i'm looking for an easier way to mask things where i'll use a brush on certain spots, Krita isin't updated to work for anima yet. Although if you have a better one i'd like to try it.
>>8887109 I assume that's the v0.37.1 one? Is it cfg distill? Because all the dmd2 ones butcher all detail/artists and this nigga stopped titling the loras with the gimmick but his main focus was the dmd2 so civjeets can gen in 4 steps or whatever so he might have just dropped the cfg ones? I'll try it out later I guess but have no reason to move on from 0.24f even on Base 1.
>>8887168 I meant that one >>8887053 which is latent mode. I didn't look for an updated attention mode one yet, didn't seem necessary. That will require a custom node.
>>8887183 i've tried that and it just won't work well, unlike normal NL/tag prompting like. >>8887116
I can do complex regionals on noob, one character with x artist, another character with a different one and even provide "holding" something on one hand and the other all in one gen, that's what I want to see in anima + regional. course we can always inpaint but i'll be better to create it as much as we can in a single setting, but I guess that's enough for this evening I gotta keep baking later.
>>8887097 mostly artist, the usual tags aged up, mature female, etc. loli in negs change her bodyshape if necessary, gremlins have too much ass while others don't have curves at all if you're changing her clothes too or taking them off, use prompt editing to set up a mature body without her tag first
What speedups can I actually apply to forge classic and neo. I usually use --cuda-malloc and --pin-shared-memory but I've wasted hours chasing other optimizations that didn't even work.
>>8887343 I am using >--xformers --flash --sage --cuda-malloc --cuda-stream --pin-shared-memory --lowvram I have a nvidia card so idk what more you can do than just that
why are you specifying 3 attention functions? IIRC there was some reason to also have xformers installed alongside sage but definitely no reason to still be using flash attention.
>>8887351 I also like to do random lines on MS paint from time to time but I am poor so I do it with my dying mouse that I bought from aliexpress a year ago :(
>>8887351 Are the two existing ones not good enough? The Lllite canny one was ass, but the cosmos reference thing seemed to work alright for canny. Lllite has better anytest and inpaint.
>>8887422 >how much money can i make with a blacked twitter account? loadsemoney just do it you need a minimum follower count to monetize though. you can jumpstart it by buying them from fiver :)
Regional test. It is kind of a pain with more than two areas, takes a while to dial in a stable prompt. Also using different artists makes the characters' relative sizes inconsistent, scheduling some slop bias on the early steps would probably solve that
>>8887476 >linux since I made the jump AMD or Nvidia? If I had the former I'd have switched years ago, but from what I can tell, the latter still has a noticeable performance overhead compared to wangblows.
>>8887479 Nvidia, the proprietary drivers are getting there from what I can tell. They're gearing up to make proper contributions since Linux is seen as the ML OS, just gotta wait.
>>8887480 Well, there's hoping. I always found it strange that Nvidia of all people, THE company dominating the AI backend market, has such poor drivers for THE OS used by the overwhelming majority of said backends.
>>8887479 >latter still has a noticeable performance overhead compared to wangblows what the fuck are you on about, literally everyone use linux for ml, nvidia squeezed everything they could out of their linux drivers for ml tasks already
Once the images start to have a "ringed" hightlight on the breasts at high epochs while barely or non in the lower ones, does it mean overbaked/overfitting? and are repeats bad in a rex/came combo as I set them to 1.
>>8887482 >the only problem was directx12/vkd3d perfomance Yeah, that's a pretty huge problem. >>8887483 I obviously meant game performance, I don't just use my GPU for AI and nothing else.
>>8887485 problem is: weight 1, max epoch: better body, has "ringed" highlights, weight 0.9: zero highlights, worse body. it's honestly odd, every epoch saved at increments of 5 have slighty different positions too.
>>8887488 >every epoch saved at increments of 5 save every epoch to make it easier for testing >came rex at this point, maybe try a different optimizer. use adamw if you want to bake low and slow. if you want to bake for an undefined number of epochs, try a schedule free optimizer like AdamWScheduleFree or ProdigyPlusScheduleFree
>>8887493 adamw overbakes quite a bit for me, correct colors, better body but zero unique positions (from side will have the same position as the sample, etc.). i'll try this prodigy next but I have no idea for it's settings + optimizer args.
>>8887495 >zero unique positions (from side will have the same position as the sample, etc.) well then increase the batch size and gradient accumulation to make the lora generalize better. try global batch size of 4 to 8. remember to scale the lr
>>8887497 i'll set it to 1 the, is it fine for RAWR to stay too?
in optimizers do I keep it betas 0.9,0.999,0.9999 or just remove that and weight decay?
>>8887499 I thought grad accumulation was bad where it's only used for the lowest of vrams, I think I cap at bat 6-8 though, (16gb vram)
Either way i'll attempt prodigy way later after i'm done with the last two came attempts, like it's almost there but i'm completely nitpicky about the quality.
>>8887495 How do we install easy training scripts in the modern day? It hasn't been updated in so long that when I install it, it says there's no torch 2.6.0
>>8887511 Scheduling is always unstable, even for concepts. If a seed resolves one step earlier than usual you could get a completely different style. You can also do alternating schedule, in which case instead of trying to hit a threshold it's just a dice roll.
anon linked some artist mixer addon >>8885806 but I can't tell if it makes things any better style loras mix fine as long as they don't have a trigger prompt
ai toolkit is gonna add anima training soon that is what I am waiting for to be honest, can't be bothered with 10 different trainers that won't even be maintained
>>8887516 >anon linked some artist mixer addon >>8885806(Dead) but I can't tell if it makes things any better Imo still unstable as heck. gaaaah I don't wanna train loras... I don't even have a good style worked out back on Illust...
>>8887518 Worked for everything except pytorch and triton, the problem seems to be that I'm running python 3.14 but I can't be bothered going down to 11/12 so guess I'll just wait for now.
Is there a good anima finetune yet Preferably one with furry knowledge I want to make a giant lizardman guarding a treasure in the forest, but I want to make him look more like a snake with legs and Anima just can't do animals for shit
>>8887579 Reforge appears to be abandoned. The thing is that Regional Prompter doesn't seem to work properly on Forge Neo. I have this sneaking suspicion Haoming02 purposefully makes it incompatible with his UI because he wants you to use Forge Couple which fucking sucks.
I am really confused on how to properly tag images for Anima. Is there a minimum of how descriptive I have to be? Should I use multiple sentences/paragraphs if the image is "complicated"?
>>8887581 Same thing with adetailer, it says that it was used in the metadata but there's no indication during generation + no difference in the before and after either
>>8887588 Haoming02 made a Forge neo compatible fork. You need to grab that version from github
>>8887589 I'm not a furry but I like doing bestiality. Regardless, your personal distaste for a fetish is a very retarded reason to be fine with something being more limited.
>>8887591 Bestiality is worse than furries >your personal distaste for a fetish is a very retarded reason to be fine with something being more limited Which is not what I said as I'm not fine with it being more limited, but you're really tempting me to think it's based just because it spites your kind
>>8887584 You should realize a lora is not large enough to re-train the model's understanding of language. Using only booru tags is completely fine.
The usual advice for tagging is, imagine you're trying to generate the image and your caption is the prompt. In fact you can try it out as well, see which prompt gets you closest. The other usual advice is, avoid any poses/compositions/concepts/etc the model would have trouble generating, except for the one thing you're training the lora for.
>>8887591 Actually, my personal taste is all that matters. I'm perfectly happy with what anima can do and glad it doesn't have anything "extra" which I wouldn't use but who knows how it would have affected what I do use, and you can go fuck yourself ^_^
>>8887591 >Regardless, your personal distaste for a fetish is a very retarded reason to be fine with something being more limited. there's a popular opinion that introducing e621 invariably causes aesthetic degradation of the models
>>8887599 >glad it doesn't have anything "extra" which I wouldn't use If does though, all of deviantart and ye-pop and clearly it's not affecting the model negatively during normal use
>>8887602 It's been that way for all the SD ones, not specifically e621 but anything you train will bias all the other styles. Anima seems really good at separating these influences.
>>8887605 I don't think noob was negatively affected by e621. They fucked up training settings which lead to some issues with noob but it didn't have the "sepia" problem which is widely believed to be the e621 disease
>>8887610 The only model that really had the Sepia problem was Pony and Pony was such an utterly shit model that it was a miracle that the thing even worked at all.
>>8887618 I still can't recreate some of my illu 0.1 mixes on later models... Too bad because they were jpg artifacted to hell (also eps ew) but otherwise I like them better than what I got from noob.
>>8887602 >there's a popular opinion that introducing e621 invariably causes aesthetic degradation of the models it's actually not just that. That it has a lot of overlap but different variations of tags actually causes degradation of the model as a whole. You can see an example of this at work by looking at noob eps 0.5(the last iteration of the eps model before TE training was enabled) and all of the subsequent eps variants, as well as how poorly the vpred models were received.
>>8887625 I like anima better, it can do full range of styles while NAI (v4.5) seems quite limited. Inpainting model is amazing though and I use it routinely to fix anima gens. Serious answer.
>>8887625 Anima not having e621 tags means there are so many baked in concepts you can't negpip away. You're essentially forced to use the very limited repertoire of danbooru tags. The natural language understanding doesn't appear to be great. Also it has anatomy problems.
>>8887638 Very useful tags like mostly offscreen male, obscured penetration , or three-quarter view tags off the top of my head. There are no similar tags in danbooru. There are plenty more tags that are very useful to include or negpip to better compose your image.
>>8887625 I have a NAI subscription and use both it and Anima (increasingly just Anima).
NAI pros vs Anima: - Knows more. Especially less common characters and artists. - Some artist styles are like 5-10% more faithful
Cons: - God fucking awful default style, with no artist tag half the gens look like someone secretly added (bad quality:1.5) to your prompt - Significantly worse NL understanding thatn Anima. I don't know how that's possible since Anima has a 0.6b text encoder but it is - Overfit to hell and back. Try to make a girl with straight black hair using artist abmayo, it turns her into a blue-haired Miku because that's how overfit the model is on his images - Worse anatomy than Anima, it's like it doesn't really understand how things are supposed to connect, it's just been trained a shit ton to brute force its way through - Bad knowledge of things not well represented in booru images. Like random objects you try to put in the image it just doesn't know how to do. NAI has been trained from scratch on only anime images and nothing else, and boy can you tell
Anima is clearly the more powerful model it's just a smaller dataset / undertrained compared to NAI so it loses on sheer breadth of anime knowledge, but wins at everything else. Oh by the way if you haven't noticed its an open source model and has hundreds, soon to be thousands of loras. Honestly it's kinda hard to justify a NAI subscription at this point and I might cancel.
>>8887674 the character knowledge is more of a dataset prep/dropout scheme sort of thing than anything else. It's also the only thing nai is actually good at, they're pretty shit at anything else.
>>8887674 >NAI has been trained from scratch on only anime images and nothing else, and boy can you tell It has full furry dataset though since it's 2 models in one with a little switch in the top left.
>>8887680 ok sure but my point still stands, NAI is anime, furry, pixiv, literally only trained on drawings. No realism, no generic internet scraped images. Anima's base model Cosmos was trained probably on ~1 billion images, and it has preserved a lot of that knowledge, it can make things that you just don't see at all in anime images.
>>8887684 I doubt it's actually only that. I wouldn't be surprised if they used COCO or something. remember, they had to retrain after v4 and then retrain again because people were finding ways of making it generate CSAM(read: photorealistic children, which does actually get classified as CSAM by the feds).
>>8887686 It does seem like anything resembling realism is heavily nerfed on 4.5, but if that was such a problem if doesn't make sense that they also left v4 which can do some really questionable stuff (no proofs).
>>8887687 isn't tile trained by giving the model a bunch of low res image with high res pairs, how they figure out anytest and inpaint but not tile? what is wrong with they?
>>8884877 >/r/ I used to request. But instead of wishing to jerk off, it became a place of anxiety as those faggot "wizards" somehow seem to recognize what I posted, and targeted, skipped and ignored my requests. They were acting like power tripping assholes doing others except just mine in my face as some sort of bullying and unexplainable spite. Now that /r/'s gone I've been freed lmao. No more anxiety and addiction to post and try my luck with new requests, as there's nowhere anymore to post. I shot a scathing message about skipping and spiteful "wizards" as a parting shot when they attempted to resurrect their request thread in /b/ but failed hard lol. I'm now a more normal human being who's moved on to healthier shit compared to being addicted to join other brain rotted anons wanting to try their luck getting their pic done to mindlessly goon thereafter while there was still /r/ previously.