>>8887708 this is some early ayakon tier shit you really don't have to prompt for high heels and gloves when she wears them by default, they'll just end up on everyone else
Realistic tags are actually kinda helpful with anatomy, background and composition, particularly if you're not going for a pure animu style Maybe tag prompting will be useful
Why is having low shift a bad thing I've seen anons talk about it in these threads but in my personal tests it looks better than having higher values Has anyone done a x/y/z plot on this?
>>8887716 For ProdigyPlusScheduleFree, delete all optimizer args except weight_decay https://github.com/LoganBooker/prodigy-plus-schedule-free/blob/main/prodigyplus/prodigy_plus_schedulefree.py It's a schedule free optimizer so just set the epoch to something like 1000 and test every epoch until you get a good epoch then stop training
>>8887734 i'll bake straight away up till 90 epochs like usual and see the findings, it's not like I can stop it and then gen something so I must see it to the end.
>>8887736 I agree The only upside is that it makes inpainting easier somehow I suppose it's a good thing to have in handy when you want to test composition really fast
>euler_a: Softer, thinner lines. Can sometimes tend towards a 2.5D look. CFG can be pushed a bit higher than other samplers without burning the image. Yet retards in here will pretend euler a is bad
>>8887746 I wonder what they meant with that part. It smooths out gradients, but you don't have gradients in the first place in actual 2D. Could it be referring to 3D stuff?
>>8887745 euler A is shit because it doesn't work with controlnets if you use them the way you're supposed to use them and it's shit for upscaling. genning straight t2i at higher res it's "fine." not really great or preferable but it's "fine."
>>8887751 imo depends on the style you're going for I know people here hate anything that isn't anime coloring, flat color, (oekaki:2.3) but euler a 6 cfg works nicely with soft shading and some artist tags
>>8887760 LTX is abysmal dogshit for anime, it has absolutely zero idea what it's supposed to look like and will make it into uncanny 2.99D aberrations within just a few frames. Use Dasiwa (WAN 2.2 shitmix) with the general NSFW LoRA if needed. Layering LoRA on top of LoRA really fucks with the details so you should only use the bare minimum, which with an NSFW-capable shitmix is just that one LoRA for when you just aren't getting enough motion.
>>8887753 >upscaling works fine too if you understood how euler a worked you'd understand that it works in a technical sense but it's pretty bad for it. >and how are you not supposed to use controlnets euler a only really "works" for controlnets if you have them on for the full generation duration. which you shouldn't do. Ideally you have controlnet end at 40-60% step count but if you do that with euler A it'll basically just erase what you got from the controlnet and do its own dumb bullshit rather than properly building off of it.
>>8887772 It does nothing on architectures other than SD1.5 (and possibly old SD1?). But some UIs fill it into metadata, Civitai asks for it when uploading pics, etc.
>>8887772 iirc it's just cumragsoft that cares. sdxl needed it flagged to -2 or something to work properly? I think. but on any a1111 based UI, clip skip does absolutely nothing unless you're using sd1.5 based models.
>>8887766 I don't know the proper explanation but in my experience it's basically where in the gen it distributes most of the noise If you set it to the lowest, it will immediately gen most of the details and the artstyle very early on, and then it will slow down as it refines it further If you set it to the max, there will be a lot of meandering noise early on and it will slowly build up the details while accelerating towards the last steps You would think this is straight-forward, but not really. Setting it at low will make the artstyle and the details look better but can fuck up your anatomy and composition and vice-versa, plus this affects how you use prompt editing greatly
>>8887783 Due to how distributing works, if you set shift to high things like [@artist A::0.8], [@artist B:0.8] will still have a relevant effect in your gen, but setting it to low will only make 0~0.3 or so actually relevant
>>8887778 -2, zero, or nothing Anything but 1 really, because that's when it tries to enable clip skip and it's not meant to be used with models other than SD1.5. Webui/Forge prevents you from being retarded and just ignores the setting no matter what you put it.
>>8887782 does cumragsoft use - as a denotation or something and not as "negative?" because yeah nai leak/subsequent models used clip skip 2, but I mean that cumsoft breaks SDXL if it's set to anything other than -2 it would break, as a sort of function of the software itself. but no, sd1.5 models used clip skip as a sort of "layer" function. There were some really funky fucked up furshit models that tried to contain specific features/datasets to individual clip skip values. >>8887791 yeah that makes sense. And yeah, a1111 ui's completely disable it for anything not sd1.5, since it doesn't work/isn't intended there to begin with so there's no point in allowing anyone to enable or use it.
>>8887774 That makes no sense. You've already pushed the gen elsewhere than the seed would have converged to, so when you cut off guidance both samplers continue operating as if you'd just started an img2img with partial denoise from that point.
>>8887795 I have a 3050 with 8gb vram lol, the average gen with 30 steps takes a min and 20 seconds or so It is most unfortunate that a x/y/z plot doesn't accurately represent what goes on with shift since you can't visually see noise throughout the gen like I can since I have live preview on https://files.catbox.moe/d73tb2.jpg Check the artist mix to see the prompt editing
Maybe I should start posting some of my slop gens here after all. Thread has so few image posts these days it might still be fun to see for some people around here.
>>8887822 Is it that surprising? Any oldfag who's been here since pre-SDXL days you would be able to recognize from their posting or gen style, it's been years.
>>8887815 This is the only way forward. Though you can't revive the thread on your own, you'd have to get other people to genpost as well. Also should probably do it in /hdg/ and end this farcical split, it's gone on long enough.
>>8887797 >both samplers continue operating as if you'd just started an img2img with partial denoise from that point and what do you think euler a does? oh right it tosses more noise on top of it and you just lost everything you forced it to try and do(since that added noise allows it to try and correct itself back to what it WANTED to do), while regular euler will be forced to build off what it already got to.
>>8887833 This is a strangely common misunderstanding of what ancestral samplers do. At each step they remove some of the remaining noise and replace it with a new random distribution. The denoised potions of the image remain untouched. The practical result of this is they converge to a different result whenever you change the step count, because that changes the noise schedule it pulls this "random" distribution from.
You can check this yourself as well, using Comfy's Advanced Ksampler. Stop the gen with a couple steps remaining and decode the result, you will see the already denoised portions do not change as it advances through further steps.
>>8887716 SchedulefreeProdigy isin't as good as came/rex combo, this one repeats certain hair waves, higher epochs shrink the head while elongating/widening the body, including adding the hollow circular highlight I don't even have on the sameples unless I increase the lora weight as high as 1.5, came/rex will just either do one of the two things amost all the bakes, I'm reaching the goldilocks zone for came/rex anyways, but doesn't necessarily mean I won't test other optimizer/training combos.
>>8887766 shift influences how quickly the model removes the noise per each sampling (or training) step, with lower shift (e.g. 1.5) you almost immediately skip to low-noise steps, meaning composition is going to be ass because you gave very little steps to it. with higher shift (e.g. 7.0) you spend more time in high-noise region. also note that shift is the only "correct" way to adjust sampling schedule of flow matching models, "beta", "karras", "exponential", "ays" or whatever were not made for flow matching and when you use them there's a ton of math fuckery going on without a good reason. >>8887769 clip skip is a hack introduced in novelai's first anime finetune of stable diffusion (they resumed from sd 1.4), basically it cuts some last layers of the text encoder before sending embeddings to the diffusion model. supposedly the model has richer embeddings to work on that way. sdxl adopted this trick for both of its dual clip text encoders from the get go iirc, and you don't need to set it manually. >>8887745 >>8887830 the only thing euler a does differently from normal euler is that it removes the noise it initially added and then back puts the same amount of different noise. sometimes this may fuck up sampling with weird sampling tricks. maybe it may fuck something up if the model was trained with, for example, optimal transport (which tries to avoid training on very very bad noise in order to make the model rely on noise more) or some other noise energy bullshit. i'm highly surprised but this anon is 100% correct >>8887840
Unlike SDXL, Anima uses the Qwen text coder, offering the significant advantage of being able to output images with the intended composition using natural language prompts (such as images with multiple people or composite scenes).
If WAI's image quality and Anima's prompt understanding are combined, it will become the standard model for the future.
>>8887893 yeah AI let's me use my coding skills to make art, a hobby I always liked but never had the time to get "good" at because I have to afford rent.
>your best was good enough no it wasn't I grinded shitty drawing for years I was never going to be good. Although it does leave me annoyed at people I am far better at doing traditional art complaining about AI stealing their work because they don't understand the tech,
bros.. how do I deal with knowing this kimchi slut will never remotely admonish my little brown existence with such unbridled contempt while draining me into a prune with her grippers?
>>8887950 So he just described the NAI inpaint trick from the pre-reference times and then highlighted two nigh unusable LoRAs extremely limited in what they can do.
>>8887971 >trick from the pre-reference times The trick is still largely relevant and wasn't replaced by char reference. For example reference works like shit with inpainting model (they even have a warning for it) + anlas cost is gay.
>>8888046 I was going to say they missed one >>>/aco/9196124 but in the time I was making this post it got nuked. oops. (also I legitimately do not actually know what the difference between that thread and asdg was at all) >>8888050 only if the content it was containing isn't banned sitewide. if it's banned sitewide, losing a containment thread doesn't matter - there's nothing to contain.
>>8888055 sounds like a dumb and worthless endeavor. also it seems that their futa thread is mostly realism so that's kind of... funny. wonder if that'll also get nuked.
>>8888058 No big dead, it's been slowly dying for like a year anyway. I would put up a gachaslut cosplay once in a while. Just feels like the site is getting shittier every other day, and this is another step.
>>8888063 Why would they ban the only hentai that looks good? Nobody wants to see primitive doodles made by unskilled losers, much less masturbate to them.
>>8888113 welcome back, i've been extensively doing different loras lately but considering your more used to gloras, (to an extent 100% of my noob loras was based on your grand design since noob 0.5.), how do you combat with overfitting? the color, body is near on spot as it can be but the moment I do "from side, all fours, etc" it'll do exactly how the body is (i.e one hand lifted up, one hand under breast) despite not adding in the prompt, the one i'm doing is the came/rawr combo lately in the past few days never had that problem.
>>8888113 I know everyone will have different scaling of images with 4chanx so I'll just screenshot mine. Do you see the grid pattern when on a zoomed out version of the image? This exclusively happens with anima. Basically when it generates any kind of grainy/noisy small texture it repeats in a patterned way and a crapton of images look like this zoomed out. Basically why I have to try my best to generate the cleanest shit possible with anima and add noise myself if I really want it.
>>8888125 on my screen only the fern one does that, your screenshot right now and the original, i'll assume the type of texture/grid stitch a real life canvas has.
>>8888125 the grid thing is absolutely something I'm aware of/have known about, and I notice it when checking the images with windows photo viewer when it's not zoomed in on a 1080p monitor. it seems like it's just an anima thing? It happens on "full background" outputs. this also isn't dataset dependent. I've seen it happen on other datasets I've trained on. It also doesn't always seem to exist, it just happens sometimes when you output at larger resolutions. It's also possible it might be associated with one of the media tags? would need to test for it I guess.
>>8888131 Yeah it's not a dataset problem because for me it often happens without any loras. Can also happen with just a single artist with no quality tags and media tags at all. Maybe bigger resolution triggers this more but not sure. It seems that for grain it just learned one pattern that gives it "optimal" score on average and then it just keeps using it for every single latent pixel?
>>8888123 This is the setting I use with glora, again very good body, color correct. However, ~90% of the seed chance the position will have the same posing, camera angle, and even sometimes missing a limb. came/rawr doesn't do that, i've tried bat 1, bat 2 and now bat 4 already.
Either way it's kinda fun baking, i've made an extensive way to test loras and grade checkmark it.
>>8888131 oh correciton, even with a white background it does show up, it's just much more faint <https://files.catbox.moe/efykdu.png> >>8888137 yeah I think it's caused entirely by anima(or possibly just being nvidia cosmos as a base) and it seems to be related to anything that's kind of textured but doesn't seem to exist much on flatter colors. Though, funny enough, this is the same lora, same seed, same prompt but with a different artist call(kiato is in the dataset for this lora) but doesn't seem to produce the gridding <https://files.catbox.moe/widlv5.png> raw output <https://files.catbox.moe/djju1p.png>
>>8888123 >>8888138 oh right sorry I kind of had to do a lot of random shit. Anyway; anima seems kind of wonky/different. It does actually get a lot more awkwardly overfit than I'd like and have been trying to do a bunch of different things to try and combat it. Upping batch size does work to a degree but it also seems to hit reproduction a lot harder and the LR's. I've been doing a lot of random experiments/testing and so far anima does actually do a really solid job of tag differentiation for styles, which you can kind of abuse to do a sort of soft form of regularization by including a dataset from a different artist at a lower repeat count. Which, for reference, dino lora I've been posting with has dino at 3 repeats and kiato at 1 repeat(you can use whatever datasets you want, though you might want to try and make them complimentary(or contradictory) in some way). beyond that I've also just been working on trying to tune down a batch 16(or rather, batch4grad4) config, which has been going... kind of mediocre.
Gonna post this here as well but what exactly do I select to train anima? Since it's not based on SDXL or anything, what's the meta with a 4080 dataset-wise and settings wise?
>>8888146 Out of all my tests on my end UNET only achieves better prompt positioning, I do bullshit test grading on things such as top-down, birds eye view, worms eye view, angles, dark, three times per type of prompt and checkmark "if it does x, has artifacts/clean, body issue, etc." AND also without the @tag activation.
I don't know about dims honestly I won't likely understand optimization or args you lora bakers do but i'll follow it either way, so since its 32 i'll keep it as is, filesize doesn't matter to me, quality and malleability as much as possible does.
Does timestep sampling matter at all? some has sigma others sigmoid, even that russel guy's confusing lora example is in a uniform type of timestep sampling.
>>8888211 >timestep sampling this is needed if you want to train on some noise levels more frequently than others or just change how you sample noise levels during training picrel shows shifted uniform and shifted logit normal distributions >>8887847 logit normal distribution focuses more on the most semantically difficult timesteps (desu that's questionable) and basically never giving the model the opportunity to learn at t=1.0 and t near 0.0 uniform overfocuses at t=1.0 and rarely learns at t near 0.0 here, t=1.0 is full noise, t=0.0 is zero noise (so it's just your image)
if you're training a character or whatever with logit normal, it may look like the model converges faster but it'll probably have funny composition/anatomy and funny little details like textures
>>8888238 You mean sigmoid (logit normal) is considered bad? ezscripts says that's the one recommeneded. so far looking at loras its sigmoid dominant with a mix of sigma, shift and only one I found in a uniform is that greg by russel to which I feel he hides something in his preview 3 lora.
>>8888242 it isn't necessarily bad, it's just different and it has the most fucking stupid flaws that nobody actually bothered to work around (or at least i think nobody did) use whatever you like more
>>8888246 I tried my best searching for more about these but considering you know more, which one is better then? I agree with you about (sigmoid) logit normal, it gives a bit of randomness on positioning/anatomy, Sigma does things better at anatomy but it has problems of the same type of positioning so i'll assume uniform is much worse.
>>8888248 you should have started by trying out uniform and logit normal first and by the way "sigma", "sigmoid", and "shift" option names are beyond retarded and have nothing to do with anything. if you want to know what are the actual distributions (because this is what this option is intended to change) you'll have to look up the code. also you can push shift value a bit higher for logit normal, say 5 or 6, it may help with anatomy
>>8888113 Well this took a couple hours but I removed it (with python and math and help from LLM of course) almost non-destructively. "Almost", because I think this kills the very highest frequencies so the texture looks a bit incomplete now, but that can probably be fixed by adding a little bit of actual random noise in place of the bad patterned noise back into the image. I'll experiment tomorrow.
>>8888264 also if you want to try uniform after all, keep shift lower instead, closer to 1.0, say 2.0. uniform kind of becomes not uniform because of shift. you'd still want some, but you have to find shift for each base distribution individually.
all of this is used in the first place because researchers found that some timesteps produce significantly higher loss than other timesteps on average despite not being all that difficult. that's when minsnr blown up, which is basically just lowering the loss on the timesteps which throw off training. then sd3 brainrot appeared with logit normal sampling and argued that weighting uniform loss is less efficient than selecting timesteps more intelligently than uniform. well, look how it turned out. now we're back where we started, using uniform without any weighting. some choose to pretend that the composition is alright.
>>8888277 i'm gonna stay with sigmoid for now, I found out bumping LR from 0.000032 upwards to 0.000042 at batch 4 made better results, but i'll still add the higher shift. (this is rex/came combo not adamw)
>>8888282 >but i'll still add the higher shift by the way, i'd expect that it'd take a bit longer to bake in wall time/epochs. with uniform it'd take even longer.
>>8888282 it's about the same speed, I set it to shift 5 and even raise the epoch from 110 to 130 ~2hours 20 mins estimate. if it goes well then i'll try shift 6 and raise it to 150 epoch including another style lora.
I've noticed on my current bake, certain complex prompts have artifacts on it while keeping the body as good as it can be, overtraining? theirs no missing fingers or hands too.
>update easyscripts >see it going on about downloading a bunch of shit, of which includes flash attention stuff >actually test it >it finally fucking works and doesn't crash out fuck yeah
>>8888315 Changing shift from 3 to 5 didint do much, possibly made the anatomy a bit wonky, it's also odd how despite having the same seed every new bake, no adjustments to the settings at all makes a entirely different type of output.
>>8888532 >it's also odd how despite having the same seed every new bake, no adjustments to the settings at all makes a entirely different type of output. you will never be able to get the same results just by keeping the same seed in sd-scripts. a lot of pytorch operations are non-deterministic.
>>8888534 >a lot of pytorch operations are non-deterministic most of them actually are. it's just that they are not necessarily used in a deterministic way.
>>8888587 Mine are pretty good and I don't have to inpaint (much). But I stole all my mixes from another bro who probably spent plenty of time discovering them.
On complex prompts I have low diversity on different seeds, compared to Illustrious. Any advices to increase it, besides reducing prompt and using wildcards? I don't use turbo, just in case
>>8888662 NTA, but the buff dude is looking at the boy, not the girl. And the fucking twink blushing like a maiden makes it worse. Whatever you were trying to do here, it just looks like a scene prior to gay sex in a public bath.
>>8888524 I mean, you can just check the tag itself to see that the training data is pretty awful, and even then it's only 900ish images tagged with it, with most concepts that work well having 2k+ images https://danbooru.donmai.us/posts?page=2&tags=wall-eyed
>>8888731 If you go through the list of eye tags (eye concepts) you'll see that most that work have thousands of images. Very few that have less that 2k work at all. https://danbooru.donmai.us/wiki_pages/tag_group%3Aeyes_tags
>>8888737 Those aren't concept tags. But I agree that anima has some issues with tags regarding eyes and other smaller details, I dunno which part of the training caused it.
>>8888737 I can literally prompt for upside down glowing swastika eyes and it'll work just fine. you just nuked the knowledge of the base model with whatever jeet slop you're using in your prompt
>>8888759 it's like a large necklace. you start by shoving one end down a girl's mouth -> through her bellay -> out her anus -> into the other girl's anus -> and out her mouth. then you connect the two ends together
This is set to Adamw, Consine with restarts and unlike Came, rawr/rex, its at a constant average of ~600 at the start and now at it's complete bake is ~600-460 still. Hopefully I can still continue this if its underbaked still.
Came/rex from what I recall, 340-200 from start to finish.
>>8888810 The bake looks kinda funky, i'll scrap it for now as I swapped from sigmoid to sigma as a test. likeness is there but the face is off. >>8888813 Yeah, I did reiterate the tags, re-edit the bodies and it got much better, but not in the latest combo. I'll swap it back to sigmoid and rebake to see a difference. but once this is complete i'll attempt that long term warmup stable decay.
Kinda a bit sad, I used an old veteran's unet only/glora combo on noob here since noob 0.5 and I never even had to try out other optimizers/network, it just works. On Anima, it still works but replicates the positions way more often. (via Glora, Adamw, consine).
>>8888816 >define properly Properly is an adverb that means doing something in a correct, suitable, or satisfactory way. It implies that an action adheres to specific standards, rules, or intended outcomes
>>8888609 >>8888606 Whats the LR for warmup stable decay? As reference on weight I considered okay: All are bat 4. Came/Rex: 0.000042 Adamw/Consine: 0.00002 Adamw/Rex: 0.000038 (just baked)
Issues/nitpicks: Came/Rex: long torso bias, fat on different bakes. (somewhat dislikes other loras of it's OWN kind but not others.) Adamw/Consine: On spot color/body, lacks variety (same position, somehow has good lora mixing with it's own or different kind.) Adamw/Rex: Expressive faces, cleanest so far, in between the previous two, Can enjoy lora mixing with others. I thought the bodies looked kinda funky, but what I didin't know it needed higher epochs and it's starting to look good, i'll slap it to overnight at 200 epochs, with perhaps one tick lower of it's LR.
>As it stands, with these fees and the current level of Anima usage, it's still incredibly uncertain whether I will even be able to recover my own out-of-pocket training cost for Anima (over $50k) much less make enough to justify continuing to do this. I thought comfyui paid for the training cost
>>8888879 I could just revive the save state this current bake which is nice. is adamw meant to be this long baking? >>8888880 i'll try it out, give a good LR number for that might hit the same spot as others.
>>8888883 >is adamw meant to be this long baking adamw trains slower than came. also some people usually use 1e-4 when using adamw with a decaying scheduler
anima trains really well with adamw constant at 2~e-05, but use an optimizer with kahan summation. i would also do test bakes at 512 resolution if you really want to waste time tuning hyperparameters
>>8888920 i'll test it out tomorrow, I found out adamw/rawr is very flexible even at 3k steps (150 epoch), thats tonights attempts on other style loras.
>>8888928 is it a good or bad thing? i'll change it right away if it's bad but i'll do three different style loras according to the current adamw/came combo though.
how would you prompt for this kind of criss-cross top? there doesn't appear to be a single tag for it, and prompting criss-cross halterneck will get you all kinds of halternecks with straps that cross
>>8888992 criss-cross halter is still probably your best bet with heavily negging other types and emphasizing shape with NL because I had the opposite problem recently in trying to avoid it
>>8889056 here is what i noticed makes people genuinely horny on /hgg/: >petite girls that are borderline loli >oppai loli >the girls either wearing outfits that emphasizes their curves or focus on highlighting certain aspects of their body to which people consider attractive or simply wearing very casual loose clothes >usage of non-/aco/ girl characters from a 90's/2000's/2010's anime or recent gacha game >the male plapping the girl has to either be the most disgusting human being or semi attractive (but not too attractive!) >multiple girls pleasing the same male >acts of sex that are more than just 1girl pov vaginal >a story can be interpreted from the image, e.g to give context to the sexy times >interesting perspectives or/and complex backgrounds as bonus
forget anything that involves: >milfs/mature females (unless paired with a shota, but it's still a gamble) >handsome males >traps/otoko no los (they have gone out of fashion) >BBC (too much ironic posting of it has desensitized people here) >anal >memes >obese/plump/chubby girls >a style that is similar to the default look of ai anime slop these will not get you any (You)'s or box requests.
>>8889083 >anal i love this >milfs/mature females i like this >plump/chubby girls i like this >a story can be interpreted from the image, e.g to give context to the sexy times a cool scenario i can play in my head is what makes me hard
>>8889124 Yes, I'm too fat to lie on my stomach but I have no issues holding a handstand for several minutes when it's a bvll having its way with my waifu
>>8889118 He is obviously lying in the same way as 1girl. Also you made 1girl erotic and not an ugly congregation of fat so it's not even the same picture anyway.
>>8889165 shouldn't those losers be on that one shitty "progressive" twitter clone that's become so irrelevant I can't even remember the name of it anymore instead of heil hitler xitter?
Also whats this [tag|tag2] type of prompt? that's new to me, a true 50/50 mix?
And mass weighting i'm used to ( tag, artist, etc:1.2) type of mass weights. rather than (tag:1.1 ,artist:1.15, etc:1.3), was I wrong all along? the prompts look fine on my method since noob, but could be anima new.
>>8889177 >[tag|tag2] there's one syntax that will apply the tags on alternating steps and another syntax that will first apply tag1 for x % of the prompt and then tag2 for y %, but i can never remember which one is which.
>>8889222 By the age Sasha-chan could get that kind of body, she probably would have already died in a freak accident involving an hydraulic press and several dildos modeled on the most famous race horses of Japan.
>>8889224 >>8889228 Isn't the author of this thing some edgelord who keeps drawing this girl getting strangled, beaten, raped, drugged and so on? I doubt she willever reach 14.
Although I heard he made completely SFW comic. I just don't know anymore.
>>8889257 He always did some SFW on the side like exhentai.org/g/775203/5e87a59fa9/ and he doesn't just draw hardcore abuse stuff. Sasha's books are mild compared to some of his nastier work, probably because she is legitimately his waifu. Also, he actually did a book where the shota gets to fuck her https://exhentai.org/g/3546198/a37dad9926/
>>8888524 I got this image completely unprompted a few days ago. It would seem it knows. I can't check your links cause you're retarded and used litter but have you tried upping the strength? Depending on the tag you can go upwards of :4.0 without frying anything.
>>8889267 my machine probably uses less than 350~w to actually gen things(absolute peak is probably under 450w but that's only under heavy CPU load). I have a window AC unit that's less than 500w unironically I use more electricity in the winter keeping the room warm than I do in the summer to keep it cool.
Some vibe coded forge neo extensions I've been using for myself, posted a couple here before but they have been iterated on it since. All have long readme's included, integrate with each other and other extensions (i.e. tag autocomplete) and so on..
https://files.catbox.moe/bhy4n1.zip >implementation of https://github.com/KeithZ117/Comfyui-anima-sampler in forge neo, also added presets for img2img upscaling passes/etc.. I recommend "Cosmos 2.5 Default — UniPC · RF linear shift5" with 35 steps between 5-7 CFG and an 0.5 denoise send to img2img upscale (no img2img upscaler selected) with "[Upscale] Detail — PC3 · S-tail shift5 · clean pass" using 6 CFG.
https://files.catbox.moe/02ok2g.zip - >Resolution picker with a shit ton of options. standard preset total ≤ 2176; high-res preset total 2240–3072 (capped 2048 per side); on send to img2img/inpaint, Resize-to is auto-set to a multiple-of-64 target whose W+H lands in 2816–3200. Every Generate click silently logs the current positive prompt into a 100-entry session ring (prompt log)
https://files.catbox.moe/05ykxs.zip >Adds a new tab for very in-depth Anima prompt building/formatting, also has a booru tag grabber more advanced than other extensions (i.e. ranbooru) that supports danbooru/gelbooru/aibooru/rule34 and integrates with prompt builder/other webui pages(extensions). Has advanced filtering to only pull artist tags between a certain post count, blacklist words, etc, etc, etc...
https://files.catbox.moe/i4gjbl.zip >Implements a new tab for JoyCaption, will auto download the desired models and has integration with other tabs/extensions. Includes a lot, will also auto downscale images to around 1024x1024 if images are sent from a booru grab or something, its pretty autistic.
>>8889283 Yeah they were both set to local, but for some reason just running run.sh in the root folder doesn't start the backend automatically too as it did in windows. Have to manually start it. Anyway, now that I can start training, what are the recommended settings for Anima anyway? Do you need to set some noise offset like Illustrious recommended or anything?
>>8889292 >Do you need to set some noise offset like Illustrious recommended never use noise offset. you can enable pyramid noise for anima but it doesn't seem to do anything for me so i don't use it
>>8889294 Do you add all three models to the anima args section or just anima base and the vae? Is training the unet only more effective for anima? Not sure where to start desu
>>8889306 add all three models to the anima args section >Is training the unet only more effective for anima? i prefer training only the unet. but you can also train the te my config https://litter.catbox.moe/yg7ez5yxub7scau1.toml lr 3e-5 to 4e-5, around 1000 - 1600 steps at batch size 2
Using char loras significantly reduces diversity across different seeds. Not sure if this can be fixed, need to train myself to test. On shitmixes diversity is even worse
>>8889336 Personally I play with lora weights way more on Anima than I ever did with any previous model even though I used to be a staunch believer in that it has to function perfectly at 1.0 or it is trash.
>>8889343 >why are you using a jeetvitai lora for a character with 1.4k dabooru posts Obviously because I need exact same screencap style, not a doujinshi based on it
>>8889336 Always the case with loras. Some conceited anons will tell you that THEIR loras don't do that, only civit loras do. ALL loras do that. Can't train a lora and not make the model noticeably worse in everything else.
>>8889347 >Then why are you using a CHARACTER lora you fucking retard Don't you realize that style includes specific facial features of a character, imbecile?
>>8889354 There is a practical solution and it is lora scheduling. Good luck though, I'm not sure it was ever possible with forge and for some reason there is a bug with anima that makes it impossible in comfy too.
>obsessed with some oldtroon anime >cumrag >Euler A >too retarded to use a lora that wasn't trained by a Manila resident or just using >bona fide Internet Though Guy Is this the first /hgg/ native pillar
>just realized forgekeks never had lora scheduling, something that I've been using for years and took for granted like artist scheduling Holy shit stone age fuckers.
>>8889398 they should tape my shrunken, sorry brown self onto the shota's penis before paizuring it, only for the tape to eventually come off. leaving me trapped and forgotten in a hot and humid hell of tit-flesh and cum (which i drown in).
>>8889418 could you add back the shota's nostrils and make the girl lick his boogers? also add in his prostate. that would make this image truly perfect :)
>>8888930 Kahan with came/rex was so-so, about the same, could be better but I removed/added new tags, edit a few pics.
i'm attempting warmup stable decay right now and see if its good, better or not. After that, I might have to ask someone how to properly set up kahan_sum on adamw/constant next, unless it's just the same as came with kahan_sum = true
Does bucket resolution affect anything for body format? It's set at 256 min and 3072 max, for noob where all the bodies are good and okay, it was set to 2048 from the old kohya gui.
>>8889449 thanks, this is the next bake test once stable is complete. (no khanon), this LR rate just feels similar to glora's adamw/consine combo but i'll try anyways.
>>8889518 contemporary le quirky shitposting I don't even actually know the source artist because apparently this is the only image of theirs I saved and I didn't fully write out their actual xitter username(oops).
>>8889521 damn, you're good. also now I understand why I only saved that one image and why the only label I had for it was sonohen(I had no faith in the artist doing more of what I wanted of them). anyway if anyone wants this piece of shit 4mb lora mediafire.com/folder/i3ekeugv3zx3u activation tag @j34 style on its own is kind of shit but it mixes on top of artist calls kind of fairly okay? depending on what they are, anyway.
>>8889526 Pretty cute, this looks useful for a style I want to go but only draws one type of girl of the same everything, always in coats but so much enviroment and theme.
>>8889450 0.00002 weighting on this adamw, constant combo didint turn out as good, honestly outside of adamw/consine glora, anything baked in adamw really loves to turn things as if it's made by hogarth, always tall, shiny, etc, torpedo breats, etc
>>8889566 nope, if oldies here say ez scripts i'll stay ez scripts... unless kohya gui comes back which is never ever, simply because I can always ask questions, get recommendations on baking, format issues and load tomls given.
>>8887847 >also note that shift is the only "correct" way to adjust sampling schedule of flow matching models, "beta", "karras", "exponential", "ays" or whatever were not made for flow matching and when you use them there's a ton of math fuckery going on without a good reason Then what should you put in scheduler?
>>8888920 that one implementation of adamw8bit with kahan that's going around is not adamw, it's using adafactor's update clipping. if you liked it you should try adafactor or stable adamw
>>8889582 if you want to avoid math fuckery? you want to use diffusers because comfy sampling backend will fuck with math regardless of what you put in it. if you just want to see to see the native schedule you should use "simple"
>>8889650 https://github.com/67372a/LoRA_Easy_Training_scripts_Backend/blob/02bb6afaceea609d736412cd3f02c21b64d3b0e5/custom_scheduler/LoraEasyCustomOptimizer/adam.py there won't be update clipping if stabilize=False right?
e621 contributed useful learning in terms of anatomy and concepts and I'm tired of pretending otherwise. >b-but it ruined my pure animu aesthetic with le sepia and high contrast! partly skill issue on part of the trainers and mostly SDXL being a dogshit architecture that wasn't completely unusable out of pure luck
>>8889666 >[@mikoyan, @honjou raita, @haku89: @linhangwei, @dishwasher1910:0.333], NTA but is this actually stable? I've tried it a bit before with limited success, but I'm not sure if it's the specific artist combo's fault.
>>8889682 there are little germs and cheese particles floating inside that pee. and what are those germs n cheese made out of you ask? me. that's right. my sorry brown self has been shrunken down into microscopic levels before being forced to undergo endless mitosis. my little peeples now have no choice but to colonize your little penis for survival
>>8889720 >>8889722 Is it going to be unusable abysmal dogshit like all IP-Adapters before it or will it be actually decent like NAI's Character Reference?
>>8889685 >but I'm not sure if it's the specific artist combo's fault Yeah prompt scheduling can really go haywire with certain artists mixes. No real explanation why but if a certain combination is giving you something low quality half of your gens, then swap out an artist, change the timing of the steps, etc. I like to end a schizo mix with an artist with a predominantly harsher style if the early steps are soft or vice versa.
>>8889743 NAI Character Reference vs the inpaint trick with Anima + LLLite ControlNet. Also your pic has a very blatant error - she's supposed to have a blade sticking out of her forehead, not an ordinary flesh/keratin oni horn. NAI got it right like 75% of the time, Anima only maybe 15%.
>>8889748 >she's supposed to have a blade sticking out of her forehead Oh sorry I couldn't tell that from the tiny portrait in the top right and your gens looked like a black horn. Still, I wouldn't bother installing your snake oil.
>>8889742 Yes, you're right this github is overfilled with AI engineered slopcode but it does implement the flow matching sampling as far as I can tell.
>>8889833 that said, you can just drop the sonohen lora on top of pretty much any other artist lora, adjust the artist lora weight down a little and it should work fairly okay? probably works a lot better for artists with flat colors/that aren't super heavily detail oriented, but I haven't really tested it too heavily. https://files.catbox.moe/9hl2pc.png
Has anyone had any success prompting for three different girls with distinct features and keeping features from blending together? I am trying to do this >the girl on the left: \(dark blue hair, pale skin, curtained hair, medium hair, parted bangs, purple eyes, head wings, bat wings, brown wings, harem outfit, revealing clothes, purple breast curtains, see-through clothes, puffy nipples, purple see-through mouth veil, purple bridal gauntlets, grabbing own breast, deep skin, large breasts, looking at viewer, puckered lips, narrowed eyes, naughty face \). >the girl on the right: \(dark red hair, double-parted bangs, blunt bangs, blunt ends, long hair, skin-covered horns, pointy ears, orange skin, revealing clothes, criss-cross halter bikini, red and white striped bikini, skindentation, covered erect nipples, mini wings, dragon wings, large breasts, looking at viewer, fang, smirk, open mouth, v-shaped eyebrows, smug \). >the girl in the middle: \(light green hair, twintails, double bun, beige eastern dragon horns, short hair, flipped hair, hair between eyes, detached ahoge, pink eyes, green eastern dragon tail, white bikini, fur-trimmed bikini, large breasts, licking penis, looking at viewer, head tilt, half-closed eyes, raised eyebrows\)
>>8889878 I've also tried straight up NL >and I've also tried straight up NL >the girl on the left has dark blue hair and curtained medium hair with parted bangs pale skin. she has purple eyes and bat head wings as she is wearing a revealing harem outfit compromised of purple see-through breast curtains with visible puffy nipples. she is wearing a purple see-through mouth veil and purple bridal gauntlets. she grabbing her own breasts resulting in deep skin. she has large breasts and is looking at viewer with puckered lipsa and narrowed eyes with a naughty face, >the girl on the right has dark long red hair with double-parted bangs and blunt bangs with blunt ends. on her head are skin-covered horns and pointy ears. her skin is orange as she is a demon girl. she is wearing revealing clothes such as a criss-cross halter bikini with red and white stripes featuring heavy skindentation and covered erect nipples. on her back are mini dragon wings. she has large breasts and she is looking at viewer with a smug smirk and v-shaped eyebrows and an open mouth with her fang visible, >the girl in the middle has light green short flipped hair in twintails and double buns with beige eastern dragon horns on her head with hair between her eyes. a detached ahoge is visible on her head. she has pink eyes and a green eastern dragon tail. she is wearing fur-trimmed white bikini containing her large breasts as she is licking the man's penis while she is looking at viewer with a head tilt and half-closed eyes and raised eyebrows, which actually ended up having more success, but it's still not quite it pic rel but I suppose it could just be too verbose for 06b
genuine question not trying to troll: how can you possibly justify continuing to pay for NAI at this point? Compared to it, Anima has: >open weights, gen privately on your own computer >you can train it >rapidly growing lora library, 250 new loras uploaded per day >better NL comprehension by a lot >more updated dataset >official Turbo lora is now at the point where it barely degrades artist styles, while giving the model a decent default style and stabilizing everything >even if you don't like turbo it's fast as fuck so good for iteratively refining prompts while experimenting, literally 3x faster than NAI >non-anime knowledge, up to and including realism if you know how to prompt it and use Z-Image or Flux Klein as a refiner
NAI has: >better rare character knowledge because they overcooked the shit out of it >maybe some artist styles are better?
Like seriously what am I missing? NAIv5 could be a huge leap forward again, but right now Anima is simply better and more flexible at almost everything.
>check out turbo gallery for shits and giggles >90% of it is futatroonery, faggotshit, furshit, /aco/slop or 3DPDgarbage Not sure what else I was expecting.
>>8889928 In the HENTAI GENERATION GENERAL thread from the taiwanese nais-weaving forum known as "the hacker known as 4chan" the no genners who complain and shitpost all the time are the subhumans. ( ͡~ ͜ʖ ͡°)
You don't need to keep shouting "no u" bro. Keep posting your ugly brownoid subhumans that are "just like you", it's important that images keep being churned out.
>>8890002 >I have a dream... that one day whole world becomes brown just like me No need to be asleep, American dream already came true, you can experience it in reality.
>>8890050 Why don't you post more fat apes of your kind. Don't forget smegma, foul smell, yellow teeth, and if you put butthole in frame - maybe add some crust from unwashed poop. It's not hentai otherwise you see.
>>8889536 > same girl, in every subject with the same clothings and prop ~30 images, absolutely no sight of skin of the girl nor even a cleavage. >Different scenes at least, but 90% involves standing, very wide scenery. >Anima just makes a nude when it's the most standard 1girl solo first ever prompt without even telling nude. So that's how it is huh? if it was noob it'll have a harder time, it's still baking so I hope it gets better, holy fuck.
>>8889878 >Make sure you tag for 3girls >assign each girl a number i.e girl1 is on the left wearing..., girl2 is in the middle wearing..., etc >use negs for any stubborn bleeds
>>8890071 wow just a like a real anime what is your workflow? I want to make sexy hentai like this for my favorite animes like Sailor Moon and Detective Conan
Settings aren't optimal yet, and I'm still testing lora scheduler, vibecoded based on a tip from an anon. You can bypass that node or try it by yourself: https://github.com/Murdunbad/lora-scheduled-comfyui
>>8890040 hey i was just testing your prompt and the missing space before the artist makes a huge difference. im not sure that first artist tag is even working until you fix that space
>>8890096 honestly i genuinely wish some people who actually like to gen and train would fiddle with this and report back, and optimize shit if possible. such tricks aren't really useful if they are only known to one autist on 4chan who doesn't even use loras picrel is a janky way you can do it in comfy right now
>>8890132 What does it do for images? I've never ever add negatives in my tags nor know how to do it. (and absolutely avoid "masterpiece, score_04, lowres, high res" stuff.)
Typically I just slap loras together, maybe 50/50 or one 50/90
>>8890140 >>8890141 what the fuck are you talking about, it doesn't really have to do anything with how you tag your images
it's just that the way loras are usually loaded with cfg is fucking stupid normally: lora_uncond + scale * (lora_cond - lora_uncond) with this >>8890132: base_uncond + scale * (lora_cond - base_uncond) makes loras a lot stronger
>>8890142 >>8890096 You said it apparently only loads positive prompts within the loras, either way I got confused on your meaning, my bad, typically I just lower or raise the values on the normal lora thing.
>>8890153 it says: #timestep_sample_method = 'uniform' In his github toml config, but I don't really care about it anymore I found a "good enough" formula.
>>8890156 https://gist.github.com/tdrussell/3f79596efb8e27672da0881afd9c1d51 This is where I checked his config, either way it's alright now, not really a necessity to pursue testings for now.
>>8890227 Can you provide us with some pointers on how to properly represent PoC pubic hair drawing from your lived experience as an avid BLACKED viewer?
man I had no idea line breaks and spaces had an effect on the output with Anima same seed, same prompt aside from adding an extra space somewhere or a line break, and you get a new image is this because of it being LLM based?
cumrag niggas really be posting 1 (one) seed as proof of whatever snakeoil they are peddling this week because they still can't do X/Y plots in their piece of shit UI
which one do you prefer the most? (pic related) i based them on previous /hgg/ suggestions for mascot design (we only got so far as to give them pink hair, drill hair, tomboy, mesugaki).
left (design 1) is very regal i feel like and represents the hgg of the past, e.g as a more formal and less shitposty version of /hdg/ - but it doesn't represent the hgg of the present really :/
the design to the right (design 2) is more akin to how i imagine geniune true hgg'ers today who help people with making better gens and loras, but maybe the design is too boring for most people tastes.
>>8890442 >if you have constructive feedback, let me know! kys. it starts with a mascot and soon we'll have a /hgg/ version of that one /g/ image with all the comfytroons
>>8890442 You've already made a mascot, don't you remember? then kept posting it with a weird giant body with wide shoulders (so a man) and probably never once wondered why that may not be received well here. Now the cycle begins again. I disagree with the other fag though. /hgg/ does need a mascot. And a discord server. And an avatar circle-jerk. Otherwise there may still be a (false) illusion that something good can still come out of this place.
>>8890471 >the aristocrat design done. her vagina is now sealed with a magic talisman. >>8890472 >not cunny enough once you get the character prompt in your hands you can make them whatever proportions you want
>>8890507 Oh wait it was in this thread so here >>8887958>>8888037 enjoy. Probably the last 2 for this general though, rest will go to /hdg/, gonna make friends with IA schizo.
>>8890597 I stopped using Euler A when I started noticing the weird hatching texture it does, but other than that it's great yeah I switched to normal Euler and it's alright. I found everything else messed with anatomy too much Like er_sde has great texturing but it's less stable anatomically
>>8890614 The breakdown is in the form of artifacts that looks like a cross-hatching pattern. I'm assuming anon saw them and mistakenly attributed them to Euler A instead.
>>8890628 As I was trying to explain what I was trying to do, it come to my mind that maybe explaning with greater detail what the composition and where the subjects should be placed was the right thing to do and it was lol, lmao even
>>8890629 mixing artists is for jeets. It's like walking into a professional kitchen and starting some unholy blend of marinara and hollandaise despite having never cooked before. tldr: artists know best and you can't improve on their innate styles.
>>8890634 Yeah, as long as you write what you want in a straightforward manner without overwhelming the tiny ass LLM with useless flowery purple prose and adjectives, anima will usually deliver what you tell it to do.
>>8890648 Anima can't blend styles like SDXL models could, as style blending was the result of how CLIP works and you can't replicate the same results with scheduling.
So I just started genning with Forge Neo UI and Anima. Generally, what are the settings to use for the upsacaler if I want to preserve the raw gen as much as possible?
>>8890701 >Upscaler 4x-UltraSharpV2_Lite, 2x-AnimeSharpV4_RCAN, 4x_NMKD-Superscale-SP_178000_G >Upscale by Depends on your base gen res. Anything from 1.3 - 2.0 can work before it starts to come apart. >Hires Steps Say your base steps are 30, then anything between 15 - 20 works. >Denoise Probably the most crucial one. 0.2 - 0.3 might fix and add more detail while keeping close to the base gen. 0.35 - 0.40 will start to show some minor changes.
>>8890650 I don't think this is true even if people keep saying it because you can mix styles on SDXL using loras and no tags just fine, you can probably do the same on anima using 4 different artist loras at low strength
>>8890766 we were talking about mixing styles with baked in artist tags. of couse mixing styles with loras still works. >I don't think this is true even if people keep saying it the man himself said it >I do agree that artist blending is different (and worse) than SDXL, but I think this was always a happy accident of how CLIP worked and that the downsides of CLIP are not worth it. >Artist mixing doesn't work the same way as SDXL, but this seems to be true for all models with LLM text encoders (it is true for NovelAI for instance).
I have seen a lot of nais :DD metadata and yes, nais :DD can't mix shit, one artist comes on top and the other 10+ artist tags just tweak it a little bit, you can't really make a full original (but dishonest) style with it.
>>8890801 https://blog.novelai.net/anlatan-acquires-hgx-h100-cluster-4b7a2e6a631e?gi=43e36623a2ca yes. THEIR cluster. They either own it outright(with it residing in a coreweave datacenter) or have a written contract that it is THEIRS for X amount of time. Either way it is THEIRS to do with whatever they want. And it's worth more now than it was three years ago.
>>8890806 absolutely fucking embarrassing that they've had unlimited access to 128xH100 for over THREE YEARS and are getting mogged by one (1) guy finetuning an nvidia robotics model with $50k
>>8890984 me? without really specify I cycle between working and doing chores on my home, playing games and using the slopmachine to entertain me (that being in form of text or images)
>>8891022 your real prompt is just >double v, looking at viewer, smile, happy, half-closed eyes, detailed background the rest are pure """aesthetic""" tags, what the fuck
local is still living in 2023, it's the equivalent of the guy telling that the automobile was useless becauise if people wanted to go faster they would just make faster horses and not the automobile, anima is like SDXL and is not a generational jump like NAI did
>>8891183 Honestly never tried the inpaint trick. How does it compare in terms of fidelity? Is it mostly limited to what is visible, as in, does it extrapolate well? Do you think it's worth the extra effort to set up the inpaint trick?
How do you guys handle the annoying prompt bleed for some of the more popular tags in anima? I'm trying to gen things like "after paizuri, cum on breasts, cum string, sex, vaginal ..." and I keep getting cum in the pussy randomly. Even the "after <tag>" seems to be misaligned since it biases the output to have the penis be fully visible.
>>8891205 Got some noblesse oblige in this thread unlike certain knobgobblers not passing down the results of the easy padding (thus saving others the $25),
If you've got a way that is as easy, if not just slightly more cumbersome, I'd be glad to hear you out.
anima has a superior control over 1girl kino and a noticeable better but still lacking way to handle 2 girls and that's what I would say it's the upper limit of the model with pure prompt, any more than that and it's a complete cluster fuck of tags bleeding all over the place you can always go hard on inpaint but come on, that's not really a point since you could do the same on SDXL overall I like it, I still can't get the same sharp results I was getting with SDXL but I think that's ok like and suscribe for more, my name is pewdiepie and I will see you next time :broFist:
>>8891253 The guy I just quoted? Who is also complaining about his gens somehow being less sharp than SDXL slop despite the much better VAE, making me think he's actually retarded?
>>8891252 Hi, psychologist specialising in neurodivergent ways of learning here! Some people are used to prompting with danbooru tags only and not everyone is OK with having to change his way of genning to make up for a model's limitation. Especially for people with asperger and autism spectrum perks, such a change can be extremely traumatic and could make them outright quit the hobby they love. So, please consider the needs of people that don't share your neurological makeup before typing such insensitive posts. Hope this helped!
>>8891259 >how NL play into this? The first line of his post >anima has a superior control over 1girl kino and a noticeable better but still lacking way to handle 2 girls and that's what I would say it's the upper limit of the model with pure prompt, any more than that and it's a complete cluster fuck of tags bleeding all over the place >but still lacking way to handle 2 girls
>>8891262 >noticeable better but still lacking way to handle 2 girls why ignore the first part? I said it is better but still somehow lacking at the same time is better and more easy since you can do it with mere prompt but not good enough to truly control with precision the 2 girls >but regional prompting I said with pure prompt, no other tool involved
>>8889289 Not bad, it could be placebo, shift does nothing, even more confusing when I tried neo as I have no idea what that setting does, also I never ever tried upscaling till now the settings made this very "blotty, melty?". (and to an extent posting an actual pic for the first time not ez script/meta graph.)
I honestly hope regional prompting comes, I tried testing everything given but no hope, those given workflows never listen to me. Theirs a limit on how regular prompting does I really want to push the high angles and stuff, this one is just Prompt/NL. But once regional comes I can hopefully do so much things with it, I guess I will most definitely try/attempt inpainting even if I like single gens with everything on the right spot without edits.
>>8891311 I trust this one the most ever since the guy posted when I complained about seams/blurry/glowing blur, etc: >>8882063 It's seamless too, you can also add a lora of your selection.
>>8891338 It can generate non canine creatures pretty well now which like to add, unlike noob, but what I do miss is the better anatomy for humans. (body shape wise, not render.), Noob is when the moment you tell it anything non human, it'll always want to add a snout, or dumb wolf furry two tone extremely high hued furs.
>>8891440 >>8891442 If you haven't noticed there is almost nobody competent enough to successfully do a large finetune. In the 1 year+ since noob's release there have been at least 7 attempts on various model architectures (depending on how you count) and all except Anima were complete failures.
community suffers from massive skill issue and there is no hope
>>8891450 every post-noob model flopped hard even when they had pretty good money
chenkin noob 20k mugen 10k OG neta like 50k netayume 30k+ newbie 70k chroma 150k (jesus) chroma radiance, klein, zeta who fucking knows
Hell even noob blew a huge amount of money (80k) to only slightly improve illustrious. So many teams, so many attempts, I don't see how this is anything other than skill issue.
>>8891450 >Even the anima money situations seems like a clusterfuck. Does it really? I thought it was pretty straightforward: Anima is the first model funded by cumrag org's new business model meant to ensure local gets a steady supply of models instead of having to rely on random furfags/horsefuckers/chinks who manage to shit out one usable model each to then never be heard from again. This was the announcement at the time of the first preview's release at least.
>When you were partying I studied The Blade >When you were having premarital sex I mastered The Blockchain >While you wasted your days at the gym in pursuit of vanity I cultivated Inner Strength >And now that the world is on fire and the barbarians are at the gate you have the audacity to come to me for help?
>>8891474 I don't see how it's worse than the previous status quo. Something like this, where the local community finally centralizes their monetary efforts instead of being an "every man for himself" Wild West where sustainability is completely out of the question, is what I've been hoping would happen since the Waifu Diffusion days. I don't expect it to get all the details perfectly right on the first try of course, but the notion of it happening at all is pretty groundbreaking.
>>8891474 Bad for Civitai since they have to pay a cut off every online gen, bad for anyone looking to build on Anima commercially. Which in theory includes Patreon pics, but who's gonna check/enforce that.
On the other hand it does make up for some of the training cost, unlike illu and noob which only got some Monopoly money for their trouble. illu made 25K worth of buzz as of writing this.
>>8891563 It's just a lora, gotta wait for civit to fix their servers for it to be downloadable. Also it just so happens to be the one with no character tags https://civitai.red/models/516432/akatsuki-no-goei-or
>>8891595 sir I am aware and I do redeem the rescaled CFG in normal circumstances but I wanted to make flat colors for this gen I hope u take no offence sir
>>8891612 Did you, the colors are within a normal level of saturation, the contrast is not particularly high and there are no egregious artifacts. Do you think that "fried" just means "flat colors"?
>>8891624 LAX really didn't want to train vpred, so something along these lines: >you kept calling eps grey, FINE, you can choke on your vivid colors, dumb freakazoids
>>8891749 Even if you set parameters aside, in a world of Nano Bananas and GPT-Images that have seamless editing and visual reasoning, it was architecturally outdated even at the time of preview1's release.
>>8891750 brute forcing trillion parameter models that need entire clusters of h200s to run and the electricity of a small city isn't technological progress tho.
>>8891759 Perhaps, but what they do prove is that the tech IS there and CAN run on an average PC. Just because they failed at making good models doesn't mean it's impossible and shouldn't be attempted in our niche.
>try to generate some figurines on anima >figure \(medium\), photo \(medium\) >get nothing but 3DPD or SD1.5esque slop How the hell do you prompt for this? These are the only two booru tags consistently used for figurines.
>>8891783 Very useful, thanks. I never used Gelbooru since they don't have a usable blacklist system so the thought of using that site's tags instead never crossed my mind.
>>8891792 it could just be because gelbooru tags never really changed and this one has a certanin cutoff right?, theirs new pointless, useless tags for danbooru like "alternate breast size", "topless female", "alternate breast size (larger),
>>8891816 nta, but I get you, I like slav squatting/squatting riding but 90% of the time its not the frontal knee type of squatting I want to see more, even with knees up.
>>8891831 change as you see fit and I am also using the anima-lllite-inpainting-v2 CN model with all of that now it feels like inpainting on SDXL again proofies of my lastestest test gen >https://files.catbox.moe/0qkkg4.png
I'm thinking of baking an oldies style of lora, but images are at a low resolution (1024 x 768), would it hurt to do an upscaler with it when it comes to baking?
i'm thinking of ranch quest, or those heavy enviroment rpg ones/similar to it.
>>8891855 it's still on paper i'm actually aiming for something like n@gi itsuki, you know the old painterly, watercolor, fantasy ones. however they all have the same low resolution. and honestly all gamecg's in panda.
what the fuck whenever I use soft inpaint it clearly works, but at the last minute it just becomes a duplicate of the base gen like I didn't even inpaint
HOW DO I USE multi diffusion with anima on c*mfyui??????? >b-but why are you using multi diffsion lil bro, that's so last ye- SHUT UP!!!!!!!! JUST TELL ME!!!!!!
>>8891852 Thanks anon, I gave it a few tests and it truly does seem to work like SDXL At least until it snap backs to the original gen because I'm apparently having the same issue as this anon >>8891879
>>8891918 Could be a Forge quirk, or at least I've never seen it happen on Comfy feels like the only difference to SDXL is it needs a little higher denoise
>>8891918 No, it seems that soft inpainting is broken, not sure how it is for that anon but every time I try to gen it works during generation but at the last second it snaps back to the original gen I did some brief researching and found a thread on reddit of people complaining about this very recently, and it seems like it could be an issue with stability matrix which is what I use because it's useful for organizing shit Could just be forge neo shitting up the bed
>>8892040 >blud, doesn't know that most of the people in these threads just like big cartoon titties or they are heckin og otakus because they like vocaloids and 2hu
>>8892380 Luckily the shitposters are lazy enough that they can't shit up /hdg/ and /hgg/ at the same time so you'll just have to switch between them every once in a while.
>>8892398 the further you push it away from realism/3d the worse it animates. local vidgen sucks balls and saas is censored to shit. vidgens are still 100% meme territory
>>8892399 I think WAN shitmixes are at about SD1.5 level. They aren't winning any awards, but at least they're recognizably hentai videos, if extremely shoddy, and not whatever the fuck LTX shits out.
>>8892414 >>8892415 it's depressing. the cyber bullying stripped all the joy of genning out of him. it got so bad that now all he posts is /e/ on a safe board. the gen itself is inoffensive and bland. good enough with melted fine details. he already knew that though.
>the gen itself is inoffensive and bland >same gen but add "penis over eyes, dark skin, (huge penis:2), testicles, penis focus, precum" >OMG CATBOX??? SO GOOD
>>8892496 being able to not rely on trannylators for games/anime/manga, but it has 0 use cases for actually making any kind of money. Japan is going down the shitter rapidly.
>>8892496 Gonna read some light novels I refuse to read in English because translations can never be perfect. Don't need to make money anymore in this life so the part about economy doesn't bother me:3
If you had to pick one artist and one artist only to use for gens for the rest of your life, who would that be? For me it's got to be Tsunako. >inb4 some butiful 2.99D tbp slopper
>>8892538 you can almost be certain that every artist with more than 2k images will be shit. Either they've been at it so long that they've lost touch and draw in some boomer style or they have some low effort style that allows them to shit out images daily
>>8892546 Tsunako is honestly an exception, every model that shits itself at old artists who've changed styles over their runtime (most notably right now Anima) nails her and only her, mostly because her style has remained basically unchanged from the beginning to now. Now I understand you were talking about the artist's actual merit and not how they're represented by imagen models, but honestly, if this level of consistency isn't indication of a unicorn who's had it figured out on day 1 and didn't NEED to change, I don't know what is. >>8892549 That was my runner-up to Tsunako but he very much suffers from the problem I described so I had to take that into consideration. If it was possible to isolate his style from ~Dracu-Riot till now, I would pick him.
>>8892576 There were at least two of us back then, perhaps three. What happens is, you gen so much stuff you get tired of genning. Although >>8885448>>>/e/3096881 it works well with vidgen which still feels kinda fresh.
Anyone got a good upscaler meant for lora baking? I want to revive old stuff but they're ~600x800, 1200,700, etc so I need the least destructive/gunk upscaler for the best accuracy.
>>8892666 kinky >>8892671 thats... expected though, fine-tuning/aesthetic tuning does that, its a trade-off between high variance&"low quality" (the base model/low variance&"high quality" (the tuned one)
>>8892813 to me it just looks like he's having a jolly old time over there, unlike you, el constipation from mexico. haven't shit in 10 days, groaning while you write another shitty retarded post
>>8892814 My hate for brownoids, faggots and other subhumans isn't shared by the mexican, he loves all that. Don't insult bro for the things he haven't done, he has enough other faults you can shit on him for.
>>8892829 mindless sex gen with zero thought, love or story behind it. learn from the masterpiecer from the other thread and learn how to give me a reason to care