Classic edition
Previous Thread: >>
>>8639665>LOCAL UIreForge: https://github.com/Panchovix/stable-diffusion-webui-reForge
Comfy: https://github.com/comfyanonymous/ComfyUI
>RESOURCESWiki: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki | https://comfyanonymous.github.io/ComfyUI_examples
Training: https://rentry.org/59xed3 | https://github.com/derrian-distro/LoRA_Easy_Training_Scripts | https://github.com/bmaltais/kohya_ss | https://github.com/Nerogar/OneTrainer
Tags: https://danbooru.donmai.us/wiki_pages/tag_groups | https://danbooru.donmai.us/related_tag
ControlNet: https://rentry.org/dummycontrolnet | https://civitai.com/models/136070
IOPaint (LamaCleaner): https://www.iopaint.com/install
Upscalers: https://openmodeldb.info
Booru: https://aibooru.online
4chanX Catbox/NAI prompt userscript: https://rentry.org/hdgcb
Illustrious-related: https://rentry.org/illustrious_loras_n_stuff
Useful Nodes/Extensions: https://rentry.org/8csaevw5
OP Template/Logo: https://rentry.org/hgg-op/edit | https://files.catbox.moe/om5a99.png
pov: you are already dead
>see some good art
>artist actually has a good number of posts
>they're almost all futa on female/futa and anal shit
>>8647847Ok what's the catch
>>8647847for me it's wokada degenerating from haha miku in a corset to chastity piss gangbang futa
>>8647847The pain of being (mostly) a vanillafag
Alright, after feeling like a retard with 291h because it sucks with hires fix (and img2img and inpainting and adetailer) I decided to do some testing
Here's an image that I did on 219h and then upscaled it, which gave me a completely different image (running into the same conundrum of the previous thread: it looks blurry and shitty if it's low denoise, and it changes the image while adding little detail if it's high denoise). Then I kept the same settings, changed the model to madlymix (the one I normally use, it's pretty analogous to WAI NSFW) and upscaled it again, and it gave me a great result.
Why is 291h like this?
>>86478631st of all 1dickgirl standing will most likely lead to warning or ban, this is /h/.
For your question, if for some reason you want to keep using 291h for upscaling even though you found another checkpoint that works: keep lowering cfg until 291h behaves like madlymix. Lowering cfg will significantly reduce the amount by which image changes in the same way lowering denoise does. I assume wai-tier shitmixes are so bad at prompt adherence that they by default act as if cfg is lower.
>>8647847which artist, asking for a firend haha
>>8647868Not a dickgirl it's Zarbon from dragon ball lol, but yeah I forgot about the amazonian rule
>I assume wai-tier shitmixes are so bad at prompt adherence that they by default act as if cfg is lowerThat's an interesting idea, I'll give it some tests
>>8647868Also, on changing checkpoints
I think I'll just buy an SSD desu I'm tired of changing checkpoints taking 3 minutes
If the switch is fast enough then it's worth it for the fixing
>>8647878reforge is shrimply shit at changing checkpoints
>>8647847Skill issue. Even tiangling can do hags. That's the beauty of AI.
>>8647863Is there a name for this outfit or does it just come with the character? I don't even know how I'd train a lora since AI would know the outfit already.
>>8647878It takes me 30 seconds on my SSD but it's not nvme. Some comfyfag said it takes him 5 seconds or some bullshit I fucking hate comfy man...
>>8647847>>8647859Vanillaniggers have 6 hundred billion options and still get mad at the rest of us for getting a few crumbs.
>>8647868Didn't work
>>8647886Wait is it really?
>>8647892Yes, it's saiyan armor, I don't even think Zarbon is recognized natively.
It tends to give you a generic vegeta style armor though, so for that gen I described it:
saiyan armor, white breastplate, pink arm warmers, dark blue leotard, dark blue thighhighs, circlet, earrings, cape
Plus it affects the style (since "saiyan" makes your character look dragon ball-ish), you can sort of circumvent that by doing (saiyan:0) armor in the tag but that will give you a way less coherent set of armor
>>8647895My motherboard isn't compatible with NVME
30 seconds is more than fine for me lol beats 3 whole minutes
That might actually make using refiner viable
>>8647893Alright. I'm going back.
>>8647899iirc forge classic is several times faster
>>8647892Just had to add "yellow shoulder pads" to this one
>>8647907Nice. I'm looking for some armor ideas for my sci-fi story and this will work perfectly.
>>8647905Last time I used it I had an awful performance on just about everything, but then again that was a long time ago and I didn't even know what I was doing
Didn't forge become an experiment or something?
>>8647911forge *classic*
it's a different project
i think it gives you like 5% worse pictures on average because of the backend being different but it's definitely slim
>artist has a nice soft animu style
>also draws his girls with a realistic chub
jishaku's truly the best lora
I don't train loras, but I came across this artist and thought it would be fun if someone trained it.
https://x.com/dokatausazukin/media
https://danbooru.donmai.us/posts?tags=dokata_(hamanan_doboku_kougyou)
https://exhentai.org/tag/artist:dkt
>>8647895he is using a nvme. it has nothing to do with the software since both use torch
Which is closer to the dataset in style? I'm tired of looking at this.
Dataset
https://files.catbox.moe/k6txkg.jpg
1
https://files.catbox.moe/w12kem.png
2
https://files.catbox.moe/wvpx99.png
3
https://files.catbox.moe/r3et8e.png
>>8648080To me they all look like they are trying to turn the picture into a generic butiful realismmix, you should try using base vpred for this kind of comparison on a style that deviates so much from what noob was trained on.
>>8648103or base illustrious, if even noob bias is too much
>>8647896They don't know the true pain of having a grand total of maybe 2-3 artists you know that do your specific kink nicely. 90% of my AI gens is trying to make more of stuff I can't find from normal artists.
>>86480802>1
>>3>>8648103If you want the Noob->Pony 2.5D style conversion these painting loras seem pretty good at a low weight.
>>8648170Worst is when artists do it but don't really get the point of the fetish
>>8648192blame the guys making requests and comissions, I don't think artists do that on their own
>>8648218i dunno, there's a few guys who only draw feet
but they draw really really ugly feet
although true, when you're drawing slop for a gazillion people you probably don't know most of the dynamics of the fetishes
>>8648221I think if they only draw one thing they probably understand the appeal well, just understand it differently than you do.
>>8648103>>86481741
https://files.catbox.moe/7o1rk4.png
2
https://files.catbox.moe/l0zjtu.png
3
https://files.catbox.moe/8y75bn.png
FRICK i had a concept idea and i didn't write it down and now i forgot
>>8648244Stylewise:
3>2
>>1Anatomy, signature and the border is a bit of a shame though.
>>8648326i am fucking about with baking
i am half convinced riz, eigaka, and toastbreeding are just alternate universe versions of each other
>>8648437riz could be a non-furry/hyper alt account, maybe
well it's not quite the same style but close
>>8648237This is probably true.
In the end people just have diverse brains and you can't apply logic or reason to them. Rationalizing fetishes is a fool's endeavor, just as is being elitist about your fetishes or thinking you have superior taste, whether you're a /d/fag or a vanilla fag.
>looking at 30-75 image artists now
>immediately more kino
>>8648549often just means their stuff got pruned for whatever reason
Sparrow has 68 pics instead of like four thousand
>>8648564oh yeah it will be worth scraping gelbooru for a bunch of them
sadly their apis is all kinds of fucked up so i'll do it with grabber if i find any
>>8648549Popular art is slop by definition, and artists who don't draw slop aren't popular enough to live on it.
>>8648633There are exceptions
Hungry Clicker comes to mind
>>8648647That's just slop, painterly
>>8648647>Hungry Clickerhankuri lora doko?
>>86486722.7K images is not enough?
Does anyone have the script to right-click for metadata? I had to get a new computer, so I don't have it anymore.
>>8648681loras always work better, but I kid anyway, I'm gonna make my own
>>8648668>Calling this slopYou wouldn't recognize kino even if it birthed you
>>8648549Man, you wouldn't believe in how many PEAK artists I've found in that image range, it always leaves me seething that I can't just tag them (and I don't know yet how to bake)
100~200 thankfully also has a bunch of good artists too though
>>8648647Lack also does some pretty good stuff
>>8648633but image count on danbooru is not a measure of popularity
some artists just don't draw much, or ask to have their paid work removed, or mods don't like them
6575
md5: 0d801f3aff06472b0cc39af8135cc4af
🔍
>>8648698i am about to look myself for baking
>>8648695ironic shitposting goes to /hdg/
I found out that fujos and fags tend to be really great artists for some reason
Here's a few that I've saved:
creamyghost
mgong520
runa_(nori31291404)
ena_(enaa97)
pigeon666
dotsuco
rei_(sanbonzakura)
jiz_(pffbq)
hataya (arguable but his art feels kinda gay lol)
All of these have given me great results when used as tags, even the ones with very few images like enaa97
>>8648711It's only partially shitposting, I do think his style and composition are great
>>8648715I like that the japs are still holding up the bishi flag
>>8648710Here's some with pitiable numbers on danbooru
tto_ja, tonotya_ and hiroaki samura
>>8648715The model learns styles faster the more they differ from the "danbooru average". That's why even some low image artists are very strong.
>guy's top pic is hanako from ks
i have found my home
the most fun thing about looking at these artists is randomly finding pics you jerked off to like 14 years ago
For me, it's fizrot. Wish the models knew him so I didn't have to use a lora. Also wish he did wider hips, though I guess you could also say that's a lora issue.
>>8648872models dont know him?
>>8648886Yeah, surprising given the number of posts and uniqueness.
>>8648887170 pics should be more than enough for such a unique style, but without a "gold account" he only shows like 20. And the partial knowledge definitely feels like a 20, all it knows is it's a pedo artist with dulled shading.
>>8648899Wasn't noob's dataset supposed to be complete? Was that just misinformation?
does this count?
https://litter.catbox.moe/wyycv70xcviey5q6.png
>>8648900Who knows for sure. But it's weaker than other artists in the 100-200 image range, and with a style that "should" train easily.
>>8648900eh based on my schizotesting it's just random chance
and also really depends on your shitmixes and even sampler/scheduler
but that's why i baked for every artist i wanted to use, not worth using inbuilt artists
>>8648914Unless they look good. Some are inaccurate in a good way.
Does someone know an artist/lora to get slim petite girls with shapes like this? https://files.catbox.moe/2i230n.png
This is peak femininity for me
>>8648916pony had the best wamuslops
but naah all the vpredshit massacred the artists, at least on what i'm using
>>8648900nta but I remember reading early on in noob's training that only angel had the complete danbooru dataset and that noob didn't have the gold stuff, but that may also only have been for the early versions
>>8648904try v3
>>8648920There are plenty of cool ones even on base noob in the bigger numbers. And a few I really like even from low image ones. Some picks:
(207) nanameda kei https://litter.catbox.moe/7q9pokq146e1bgsh.png
(158) tenk https://litter.catbox.moe/jmko2lbxhr0g963h.png
(87) kcccc https://litter.catbox.moe/ct45d816ugd451cb.png
>>8648927Might be. And he only has one image on e621, which is what helped many other artists.
>bake style lora
>it overfits on the poses
many such cases unfortunately
>>8648938On comfy I just [<lora>:0.3] and problem solved:3
>>8648918white datura and as109 do hebes decently
>>8648918try rustle? he's one of those uncomfortably realistic ones
>>8648944>only activating the lora after the first 30% of stepswhy didnt *I* think of that...
>>8648944Care to post a workflow? I tried scheduling loras using the PC schedule loras node but it seems to have a bug that doesn't allow it to schedule at the right steps, on my machine.
>>8648970It's just prompt control anon, you really should have it. It has all you could ever possibly need from comfy++ emphasis (which is way better than default comfy and a1111) to lora scheduling. It has example workflows in the github too.
https://danbooru.donmai.us/posts/8928950
>large breasts
https://danbooru.donmai.us/posts/9569141
>huge breasts
Why are the people tagging danbooru pics so fucking retarded
>>8648326i really like your oc bwo
>>8648970As a workaround you can use the refiner flow and just use the same model for both, only with different prompts/loras/etc. Does the exact same thing in practice.
>>8648979because (you) didn't fix it
joke aside i like that there isnt even a tag for large ass
fucking loathe danbooru
>>8648979I wish people tagged them to exhaustion
Having more tags in an image, so long as they're properly tagged, sounds objectively better
it barely does anything but fiz-rot at least does something compared to fizrotart when prompted
these smaller artists are definitely nice for baking
all quick
>>8649002give overfit-proof toml
>>8648979And I want to go even more precise personally. Ideally we'd be able to prompt using numbers to specify the size of each part of the body. There are ways to make a dataset to do that if only they had the foresight,skill, time, money, etc to do it. Same for stuff like camera angle.
But in the end I feel like blender or some other 3D program with controlnet is probably the only solution for truly exact anatomical precision. Humans can't simulate all of physics in our heads and we can't expect a weak AI to do it either after all. It'd be nice if there was a package you could download with a body that features muscle and fat simulation so you could get something truly accurate. Maybe it exists, not sure.
test
md5: 9a0df4f6ac2f64f2cd80d6056e4e8def
🔍
>>8648979It's not a huge issue though. Some of the "moderate" sizes work relative to her default rather than an absolute measure, if she's a named character. But it would anyways be relative to her bodytype, height, age, clothes, etc and of course style, all due to how tag biases work.
chars
https://danbooru.donmai.us/posts/9567492
https://danbooru.donmai.us/posts/9472650
>>8649001I tried fizrotart, fizrot, fiz-rot; as well as prefixing them all with "artist:". Even together they're far weaker than any one of
>>8648930 just prompted normally.
>>8648979Is she the first vtuber to focus on breast expansion as a gimmick?
>>8649009yeah but you don't always want a huge ass
you want a tastefully large ass
it still works through nl or whatever but yeah
same goes with all the retarded hands on x tags
>>8649010It at least produced something cunny shaped with a basic prompt of "fiz-rot, 1girl" whereas fizrotart produced a 1girl looking like something out of a semi-realistic slopmerge from civit
Currently working on another lora but I like the style so I'll try my hand at it later. Though if anyone happens to have a dataset for him ready, try training the tag fiz-rot and not fizrotart
>>8649024Also big butt from e621, they have four sizes. I haven't tested if they give more control.
>>8649003>>8633915not even overfit proof as much as it needs the same amount of epochs to converge regardless of imageset size
just tested one that fully converged at 250 steps lol
>>8649026semi-realistic slopmerge is noob's default style
well it's actually more like a cheap 3D render
fiz-rot is the e621 tag, there's only one image now but might have been more before and got pruned for loli. The loras I saw for him on civit didn't use a trigger tag at all.
>>8649031noob's default style is a flat red blob, semirealistic slopmerge is semirealistic slopmerge's default style
>>8649031His aliases according to danbooru are fizrotart, LLAMACOW, LLAMAvsCOW, MoeMortem and fiz-rot. If I remember correctly danbooru changed his default tag to fizrotart (from fiz-rot) sometime after noob started training
man i ought to bake on these old ass mugen animations
test
md5: bdf166a12f6774f5b6373327d602b196
🔍
>>8649033That's if you're using constant CFG like an epsfag. We already figured that out with EasyFluff 20+ months ago, and it's what the v-pred paper addresses as well.
This is a batch 9 of "1girl, standing".
>>8648927looks weaker i think?
https://litter.catbox.moe/ugszn6wuxffz4gr0.png
christ i need to finally make a differential thingy for removing tags from datasets
i always remove the same few anyway
>>8649073alternate x
futanari
genderswap
character tags
twitter username
stuff like that
>>8649074i hope you mean you remove images with these tags present and not just the tags
>>8649075No? These are pretty much always mistags. I'm not making futanari datasets lmao.
>>8649076ive never had a futanari mistag lmao
what tagger are you using
>and that noob didn't have the gold stuff
how are you retards still breathing
https://huggingface.co/datasets/Laxhar/noob-wiki
Untitled
md5: 4c0b42b4bf8695458851933548344d62
🔍
Conflicted feelings on the Millais lora
Left is the original painting
>>8649090Only that much you can put in a 4ch vae with a style that's different on half the images desu.
>>8649091You're probably right
Still a great addition to my oil painting loras collection though, since I can use it in lower weight with others
I'll test the Sargent one in a minute
>futa on male
>futa with male
>but there's no large ass tag
fucking danbooru
>>8649091cant you mostly remedy this by including close-up crops of faces and etc
>>8648979why don't you create an account to fix this?
>>8649095that always felt like quackery to me desu
i do wonder if higher base res would do anything though, it was useless for characters when i tested but i could see some theoretical merit here, these are like 20k by 20k px paintings sometimes
>>8649097try throwing it through at 1536x1536 or even 2048x2048 to see if theres a benefit
you know seeing how easy setting this shit up is i wonder why i never saved myself the time and effort of cleaning up datasets before
i'm pretty sure you can set it up in the tagger for exclusion too via a preset or something
oh well autism is a weird thing
What checkpoints are people using these days? I haven't been in the game since pony was all the rage.
>>8649102Pony is still the best
>>8649103Seriously? That's surprising to me.
>>8649106yes. it's the coldest and longest winter.
>>8649103Low quality bait
Someone please give the guy a proper response
this wouldnt happen if we kept the SOTA (Noob-AI XL) in the op
Okay, so... there are literally no new models????
Nope. Nothing new has released since beginning of the year 2024
Is Epsilon pred 1.1 any good?
>>8649117its the best of the eps models imo
>>8649120How does it compare to vpred?
>>8649122better than vpred
>>8649091If butiful saarealism like https://civitai.com/images/77496415 is possible, it should also be possible to bake a little more of the oil texture than that example image I think.
>>8649125Cool. Does it share the same prompt adherence?
And also, any refined models or merges?
>>8649126What I've noticed is that vpred is worse for this than baking on base illu. Base illu configs get more of the style than vpred configs, even when you gen on vpred (shitmixes).
>>8649131because vpred 1.0 is on the verge of exploding while illu 0.1 is undercooked as fuck
>>8649132>pwease weave my heckin lora sanctuary nyo
>>8649131I feel like I prefer eps loras (especially trained on noob eps rather than illu 0.1) on vpred anyway. Like that popular 748cm lora got rebaked by the same guy for vpred and it completely blows while eps version works on vpred very well.
in light of recent posts, i am requesting your strongest noob EPS lora toml.
>>8649139Glora config by oekaki schizo.........
>760 undesirable general tags + 4800 copyright and character bullshit
well that was worth it
any decent AI hentai animation tools nowadays? im talking just straight inputting an image and letting the ai generate a few frames beyond that, nothing complex
>>8649102It's either base noob vpred 1.0 with some style loras to stabilize it or any shitmix based on that one like r3mix, 102d custom or 291h
>>8649111Perhaps you are right, maybe I should add the popular options to the OP
>>8648981Yup, I don't like bloating my workflow if I have a choice tho.
>>8648974What do you mean? As far as I know, there is only one extension called prompt control, which gives the node I mentioned and is used in the github's example workflow. And I did try it but it didn't work. I don't know if it's a bug or something on my end. To be specific, when I compare it to the method above where you manually apply loras, the PC version doesn't schedule it at the exact steps you want. That's also true comparing it to PC's ability to schedule regular tags. For instance, if I prompt [fiz-rot:0.9], the final image is not changed much, because of course it can't change much when most of the image is denoised alerady. However, if I do [<lora:fiz-rot:1>:0.9], the final image has a lot of differences. Something is clearly wrong with the extension or my machine, even though I tried the example workflow.
>>8649153man it works
i can upload the script with the file tomorrow if anybody wants it
>>8649163sowwy i am jorking it in my bed rn
>>8649168No, but you made me check their xitter and they give out 30 free gens to new users now, hm
>>8649159There's no point sharing my full workflow, it has about 15 nodes I wrote myself so it's not going to run on your machine...
But I made a short test workflow and it works for me:
https://litter.catbox.moe/xp3yprhg9jbp5n42.png (lora switch at 0.9)
https://litter.catbox.moe/6inv4ne10rvz003n.png (no lora)
As you can see, it doesn't change an image in any significant way when lora is loaded at 0.9. Also you can tell by the little slowdown that it loads lora at step 27 out of 30.
Also sorry if you don't have inspire pack. I switched to that ksampler for... reasons. It should probably work with normal ksampler the same way though?
Bros
Does this work? https://civitai.com/models/99619/control-lora-collection
>>8649137Yes I noticed strange things like that too on some loras. Glad I'm not a schizo.
>>8649140how do i use glora in easy scripts doe
>>8649175Just... try it out and report back?
anything good lately (local)?
>>8649173Hmm, ok so actually I guess it is working on my machine, but its effect on the image is way stronger than scheduling tags. Like if you schedule an artist tag at 0.9, it has a lot less difference. So maybe it's my understanding of how loras affect the denoising process that is lacking.
>>8649158Thanks for the genuine answer m8.
is it worth trying to animate shit with WAN or is it too much of a pain in the ass?
>>8649217its not bad but nothing crazy either. it's still in gimmick territory. once we get shit on par with kling/veo itll get real, though
>>8649218okay then i wont bother, i was gonna try and run it on runpod but i foresee a ton of stupid issues and probably long gen times, ty
How can I reduce the drifting noise and instability that often occurs during the early steps with noobai?
>>8649227retrain the model
>>8649231if he was he would have done it himself
>>8649231that's quite literally the only way to fix it? idk what you were expecting. there's no magical miracle snake oil
>>8649234I know that you're autistic and thus any attempt at communicating with any more subtlety than dropping a brick on your thick forehead doesn't work, but when people ask for stuff like this they mean to ask for tips and tricks and things to go around said problem
If someone isn't capable of figuring that out by themselves, why would they be able to retrain the model
Retard
>>8649238holy moly, mommy yell at you today or something? LOL
>decided to stick to artists noob already knows
>no more baking and rebaking
>darkness works easily again
I'm finally free...
>tfw there are concepts I need loras for so I can never be free of their influence on colors, even if I can solve their influence on style by prompting and other loras (lol)
ACK
>lora i baked from an artist who only draws 1girls happens to only be good at 1girls
sad
>>8648336i should get back to baking too... but i'm still busy playing with my umas
>>8648980thanks bwo
>>8649000all my posts have stealth meta
>reforge dead
did you guys all go back to normal forge?
>>8649344nope, still on reforge
haven't seen a feature worthwhile enough to jump ship
>>8649344I haven't move from reforge, 0 real reasons to do so
>>8649094big_butt,huge_ass
>>8648918for me, it's sabotenman
I FUCKING HATE OBJECT INTERACTION AND PERMANENCE
huge ass, hyper ass, big butt, bubble butt, hyper bottom heavy, wide hips, thick thighs, plump,
detailed anus, y anus, big anus, anal hair, anal fluid, pink anus
PonyV7, Gen these prompts
instructions unclear, genned gpo ces instead
You're all a bunch of NAIggers
>>8649173>it has about 15 nodes, bro, I wrote them myself, bro, you gotta believe meOh shut up you liar, you download pre-made shit from Civitai just like the rest of us.
Enough with this tomfoolery and quit pretending you're gifted with technology.
Give us the workflow
for me, it's all styles self trained.
>pit epschizo's config against my prodigy config
>prodigy wins
heh
yeah. i don't see any proof to those claims either...
>>8649163https://files.catbox.moe/47a1mb.txt
https://files.catbox.moe/p6nmdu.py
>>8649475well ig it would be better with a folder prompt instead of using the base folder
https://files.catbox.moe/igkao9.py
>>8649475what criteria did you use to select the undesired tags
>>8649488he didn't desire them
also if anyone wants to customize it it's pretty fast all things considered
600+ image general tags is only like 9k in total with half being those repetitions like "black x"
i think it only took me 15 mins to go through them
>>8649153>copyrightHuh? Do you guys really include copyright tags into your captions?
>>8649521I think he's listing out the "undesirable" stuff to prune from dataset tags. Wonder what "character bullshit" is.
Is this thread genuinely run by AIs?
>>8649523Yeah, but that assumes you include copyright tags in the first place, which makes no sense to me.
>>8649524Because one person misunderstood a post?
>>8649525Anon, I really hope you are not baking on scraped danbooru tags. Tell me you don't.
>>8649525Ah you mean have grabber filter those out while scraping? But I don't think auto-taggers have an option to leave those out.
>>8649527Of course not, I use a custom tag pruner to remove bullshit like "bow, blue bow, bowtie, blue bowtie".
>>8649528I use SD-Tag-Editor with eva and it doesn't use copyright tags by default.
>train lora on noob eps 1.1
>looks better on vpred 1.0 than one natively trained for vpred
hm
>>8649563Can you imagine how disgusting it'd be to just get the dickhead meat in your mouth?
Nice gen tho
>>8649572I'd rather not think about that lol
>really nice artist>20 imagessigh
i will still try baking it
>>8649576i just feel for the poor girls with these severely oversized dicks
lol
>>8649468this but only for characters
>>8648920Pretty sure they use AI, also they make lolis...
>>8649024huge ass on positive
ass focus on negative
In my experience it's ass focus that gives them the silicone looking ass
>>8649383Alright chat I gave up, I think the slopmachine can only take you so far, time to properly learn PS I guess
>i MUST use the shitty proprietary tool
yeah not surprised
did hdg die or something? why are these retards here?
>>8649730somebody made like 5 threads yesterday
a thread just flew over my house.
36 done
14 more to bake
uegh
i sure hope you're not baking another nyalia lora #9 with extra grease vpred.
there's already more than 20 different nyalia bakes out there (searching on civitai you will see 13 but some of the listings have multiple bakes and that's without counting the ones from /hdg/ rentry)
if you want to be #9 it's over for you before it even began
What is a nyalia even
thought it was just some dude's civitai username
https://civitai.com/models/950531?modelVersionId=1882934
Thoughts over noobai
>he doesn't have his own opinions
tsk
>>8649797a pixiv ai "artist"
https://www.pixiv.net/en/users/8786609
>>8649798oh the vpred version is finally out? I can give it a try
>>8649798it's more unstable, has worse faithfulness to artist tags, has the opposite problem of noob vpred when it comes to colors -- they're too fucking muted. It also is more slopped than shitmixes like wai, it tries very hard to make even 2d styles look 2.5d or have some embossed style of drawing.
It's also trained on AI artists :
https://civitai.com/images/81195842
>by E 0x0194 type1I recognize the name because it's one of those that are part of bakariso's ai lora collection.
For a base model to be trained on AI art is downright retarded.
>>8649802oh
style seems all over the place
how the hell do you actually properly finetune noob?
last few times i tried with ezscripts it didn't even launch
>>8649821you can do a fine-tune with ezscripts?
>>8649823well apparently not
but the option is there
>>8649823nta but assuming that's easy-training-scripts and not something else, yeah. The last lora type option is "Full". But I have no idea what kinds of args it needs in regards to dim size, alpha, etc.
>>8649827>The last lora type option is "Full"This is a full-rank lora, not a finetune.
>>8649827that's not a fine-tune
>>8649821I personally use modified version of machina's sd-scripts.
>>8649829>>8649830well that certainly explains things
>>8649833care to share your setup? i doubt i'll have the patience but i'd still like to take a look
>>8649835sure. but take a note that I don't have enought vram for mixed precision training, so I traing in full reduced precision mode, and bf16 for some fucking reason fucks up the weights, therefore fp16.
also I've modified the training code for a bit, like random shuffling, INTER_AREA scaling for buckets and other stuff.
https://files.catbox.moe/59e80f.ps1
>>8649821I wouldn't bother with finetuning if your end goal is a fairly standard style lora or whatever. I recently moved back to making basic loras despite them taking 2x the time to bake compared to finetuning for some godforsaken reason.
T. baked a thousand finetune-extracts
>>8649849what's your config now? for me, finetune extracts are generally better quality (and training is faster, yeah)
>>8649704Just use Kontext.
>>8649842Ps1 is only for Winblows, right?
>>8649849Since I have 20k+ quality style pics I was thinking of doing some tests for fun
>>8649867That's just a script that launches accelerate, you can port it 1:1 to bash or whatever you use.
>>8649071thanks for testing btw, if you're still up for it would you mind testing artist:fiz-rot vs artist:fizrotart? interested in seeing if one is stronger than the other
>>8649850Currently on this "kind of sort of okay" config that's still underbaked for my tastes, I only switched like a week ago. Still iterating on a better config and testing some code adjustments since the wavelet implementation seems to have some fuckery in it like ll_level_threshold working the wrong way.
Needs around 18-24gb as it is, probably a lot less with bf16 / reducing immiscible noise / disabling float64 / reducing edm2 channels and all that crap.
https://files.catbox.moe/3kx4ms.toml
is onetrainer still better than kohya for training?
>>8649875>waveletHuh, is it even worth it?
>>8649842>3e-5I assume for tunes you'd generally want those kinda of lower LRs compared to LoRAs because of the de facto rank, yeah?
>>8649901You may need an even lower LR if you train in full precision or bf16+SR, like 5e-6 or even 1e-6.
>>8649891The difference is rather dramatic, you capture scratchy linework texturing and backgrounds much better. I had a comparison image at some point but I lost it.
>>8649891I personally didn't like it but that was admittedly tested on characters not styles
>>8649905I’ll give it a shot, thanks! By the way, I noticed that the y-limit for the EDM2 graph is set to 99. So, it can reach values between 80 and 100 for you as well?
>>8649907Yeah, it can get really high especially with edm2_loss_mm but the results arent super different compared to the default gemini implementation that gives weights in 5-10 range or something. If I remember right some of the enabled snake oils are the root cause for huge edm2 weights.
>>8649910Yeah, I tested it on the finetune, and disabling immiscible with sangoi loss fixed that.
>>8649875which trainer are you using?
>>8649704Still thinking about blender myself. One thing that would make things go faster is a workflow where you can pose things in VR. Posing things with a mouse and keyboard is pain.
43
md5: bbe23a5e5935b2fd14facaf56935c839
🔍
this cost me sweat (and sweat)
now for testing
>>8650078I do not recognize any of these. Which means they are bad artists.
546
md5: d90a36f02f9efc36c21954accc9c10ea
🔍
>>8650081oh sorry i will delete them :(
>>8650078ummmmmm can i have the chiisan one pls
>>8649972https://github.com/67372a/LoRA_Easy_Training_Scripts
>>8650103do they support finetuning now?
>>8649034I think you're right. Just ran this
>>8650113 with (artist:fiz-rot:1.4) on a whim, and it did something. The resemblance is there, just doesn't go far enough.
>>8650107Not on the UI afaik. You can still run it directly from the backend folder but I dont think the guy really maintains the finetuning side that much.
>>8649798>check through preview images>find some with "by firedotinc" taggedoh no
>one style goes from pixel art to smooth back to pixelart
i will never get loss
Anyone feel like image quality and prompt adherence is just a ton better when you can fit a prompt into the 75 token limit without spilling over? I hate knowing this. I have tons of complicated images with tons of things in them so I can't avoid it.
>>8650154I am pretty sure the "75 token limit" was something only in 1.5
>>8650154The positional embeddings restart when you go past one window, so the model will see all tokens in the first positions of each window muddled together, then all tokens in the second positions and so on.
>>8650155SDXL still uses CLIP.
>have OC
>need to spend a billion million tokens in order to get all the parts of the design in the gen
Haha...
I refuse to learn how to train loras.
>>8650154I have noticed this but with 150 token limit, after that my mixes lose power
>There is a significant difference between large breasts and huge breasts tag wise
Fuck
>>8650161I feel you
On this topic, anyone know how to get a tiger pelt around waist consistently?
I've tried lots of methods but it always feels like rng
>lora gens perfect artist name in the corner
heh
>>8650078For me? None of those. ;)
>>8649798you can't stop shilling your slop here can't you
How do I merge a lora and a checkpoint again?
>>8650238Baby nyalia me one more time!
can we make some posts in hdg to pretend it's alive so they go away
>>8650191Have you tried seeing if any character have this and then do a \(cosplay\) hack?
when are we getting a model trained with tags like this?
https://rentry.co/goxrrvh9
https://rentry.co/9wranqty
>>8650252hopefully never jesus christ
>>8650255you don't want to use json in your coom prompts?
ridiculous how are you supposed to make butiful hentai otherwise?????
>>8650256nah fuck that shit
gimme full html with css support so I can define divs and layer it all out exactly how I want it.
>>8650240didn't work
I even tried interacting with cathag
>>8649175The only thing this shit seems to do on my end is push the base model to the EPS side
>>8649175ZAMN I LOVE THIS LORA
https://civitai.com/images/86100517
Wtf is the difference between Reference Strength and Information Extracted settings in Vibe Transfer?
Even whoever wrote the documentation doesn't seem to know.
>>8649611true
prodigy is unrivaled for characters
Looking for an honest answer: is there a good way to do multi-character prompting on local? Regional prompter seems shit, Forge Couple seems shit. Is this just an impossible task right now with SDXL models?
>>8650425If the characters are named and known it's more or less easy to pull off with a mere prompt but if you want to put your OCs together you are more than likely to inpaint to hell and back each one of them
>>8650428I was looking more for using two character loras and being able to prompt different outfits on them.
I was browsing danbooru for inspo again and my god, I was reminded how the old spider-gwen is so much better than the modern whore looking girl that took her place.
>browsing danbooru
>for wesern capeshit
????
>>8650432Who are you quoting?
>>8650428>look at the picWait a second...
>>8650252Nostalgia
People coped about early local LLMs with that
>>8650252>minthybasis garbonyo
test
md5: 990a4e525f540443eb12f207860300e6
🔍
>>8650425Simple compositions work fine for me (on comfy).
>>8649936For more complex stuff it works when combined with regional prompter. Gen the image with mixed up features first, then re-generate the image using this one as a weak controlnet reference to fix their locations, use regional prompter at the same time, and ideally same seed. Better than inpainting them individually because this way they can interact, they get the same shading, etc.
You can also draw your own controlnet reference, no art skills required.
>>8650468mind sharing a catbox for the dual-style oral gen? that's kind of what i was looking for
>>8650431She does fit much better into not-Overwatch, but "fitting in" is a weird design goal to have.
>>8650431Both are BBC pilled, at least.
>>8650476Sure https://litter.catbox.moe/suoqwc10wni5uyfq.png
The mask stuff up top is just two rectangles, what Forge Couple does by default.
And the negs are unnecessary, it's just an old test prompt from illustrious days.
why do some people think it's okay to post/discuss westernslop here?
>>8650485The mere mention of a western character offends you? I agree it's offtopic without an anime-styled gen of her attached, but w/e. Not like it's taking over the whole thread, and we had nothing better to talk about.
>>8650481Thanks, this works decent enough. Do you know if it's possible to use regional loras?
>>8649798i compared it to noob and illust and even with its fancy prompting scheme, it still looks worse than both
>>8650481im disassociating
>>8650497Made one before with anime, /aco/ and photoreal in one, but that required inpainting with different checkpoints. If one knows realism it automatically starts to suck at anime.
>>8650491I read somewhere that Forge Couple should support regional lora scheduling by default, does it not? There's also this on comfy, but I haven't tested whether it breaks attention couple:
https://blog.comfy.org/i/152863132/masking
>>8650505I wonder how complicated it would be to get lora styles to apply to different regions. Maybe characters could work for since they use activation tags but I will have to test it. I'm using Comfy too as I had no luck really with Forge Couple the past 6 times I've tried it, but I might go back and test it to see.
>>8650487it's more up to the individual isn't it
some people just can't contain themselves
>>8650486not your personal blog
>and we had nothing better to talk about.no one forces you
cgs
md5: 9aba3eda9b5c7dec0e2b9ece77c7e5a8
🔍
Anyone want to train a Blue Archive style lora?
The PC release has all the CGs in .jpg
>>8650524upload it somewhere? might give it a try
>>8650524thats like the most generic style
>>8650525It's 400 mb
I'll see if I can put it somewhere
>>8650530mega, pixeldrain, pretty sure you can upload something like 800mb at litter too
>>8650531catbox seems down so I'm putting it on mega
It's up
https://mega.nz/file/GVJiySRJ#s33gT8t0wG1l4SAxwuZaJwcI_tuTn_Sw7XIlrofndbo
Blue Archive CGs
>>8650511Many style loras also have activation tags.
>>8650526most generic anime style >>>>>>>>>>>> generic aislop style
>>8650512Doesn't get much more japanese than melon22, he even draws for the Fate gacha.
>>8650548what is generic aislop style
>>8650550wai, autismmix, aom2
shiny 2.5d "tasteful hint of realism" garbage with ugly smooth shading
>kohya_ss just fucking crashing every time i try to click a save button
based desu
>>8650549can't seem much of him through wamu though
>>8650549i'm not even surprised
genuinely curious if you're enjoying groidspam, it's sexy henti with weaboo weman after all
>>8650573Projecting a little too hard there.
holy fuck i hate toml and python
just stop being a retard lmoa
but python was built for retards, which is why it sucks
if python is for retards, why am I using it, eh?
>an hour to try to fix retarded toml errors because the jap writing sd_scripts just put a bunch of random shit and random incompatible instructions everywhere
holy fuck i hate foss~~~!
>>8650468Are you using those image masks with latent attention or as inpainting?
>>8650589python is okay, but toml is the worst
>>8650603as opposed to yaml?
>lstat: path should be string, bytes or os.PathLike, not NoneType
kill yourself :3
>>8650605>had to specify it outside of the toml>but it refuses to load properlynice
how are you supposed to specify an xl model for finetuning again?
>>8650604yaml is better than toml
>>8650610ah yes, the language where you want to use the country code for norway and it gives you a boolean
>config just didn't work at all, had to use the command
>actually u gotta use the different script that is never ever mentioned anywhere on the site
>oh also lowram
haheha
what was the tag for that artist with the super excessive big block watermark?
>>8650620nah, much more excessive than that, it was like a big rectangle, maybe pinkish in color?
>>8650602The masks determine prompt regions, for regional prompter.
>random artist obsessed with waifu
>like 50 good quality pics of her from like the last year
Neat.
>>8650551I am a personal enemy of generic aislop style
i didn't want to use adam8bit but oh well, can get a finetune going
and anon wasn't wrong that it's surprisingly fast compared to loras
>>8650705Well it goes through steps quickly, but doesn't it need a crapton of them to actually learn anything?
>>8650709I think it's more that due to the effective rank (aka all of it and all of the layers) you just need to go low and slow to not burn it immediately
>>8650673mfw I was going to troll but unfortunately the current /slop/ thread is at it's infancy
>>8650713finally some good fucking anime.
i don't think the wd14 tagger actually does proper fractions for the min threshold
i just hope i'll have enough vram to cache all these latents
>>8649820Nice. Is there any special tag for the skin or is that just the artist?
23435
md5: da86040df1a5033136ca3d1a73054729
🔍
>>8650762I did, making a test first. I'll try letting it go for like 16h tomorrow if anything usable comes out of it.
Also, what do betas do? Do I need to specify them for Adam? I baked a bunch of loras on AdamW without the betas and nothing broke...
>>8650790it's banana_oekaki
>>8650238Your trainer is supposed to have this option.
>>8650792What optimizer gives you 1.1s/it?
>>8650796Oh nice, thank you. Should work great with Egami.
>>8650801excuuuuuse me that's 1.1it/s
lol
but adamw8bit
>>8650792>Also, what do betas do?they change the impact of momentum
>Do I need to specify them for Adam?No, if you don't specify them it'll just use default values
>>8650244Not yet, but it's only a part of the design, not the entire thing so.
I've managed to get the rest of the design working pretty consistently (despite being a very complex design), but this one part sucks
>first samples all melty
i'll chuck it up to what other anons said about samples being shitty for now
>>8650735On this topic, let's see how the "shitmix" I'm used to does in comparison with the recommended models
Oh.
>>8650812Yep. Now try a dozen other mixes and see which one's innate style seeps through them more.
hmmm i don't think this was supposed to look like this :-D
>>8650854you have to fix the state dict to enable vpred+ztsnr, sd_scripts suck
>>8650862oh for christ's sake
how do i even do that?
>>8650869from safetensors.torch import load_file, save_file
import torch
model = "/path/to/your/tune.safetensors"
state_dict = load_file(model)
state_dict['v_pred'] = torch.tensor([])
state_dict['ztsnr'] = torch.tensor([])
name, ext = model.split('.')
save_file(state_dict, f"{name}_fix.{ext}")
>>8650874where? sdxl_train?
>>8650875Just create a new script, fix the model path and run it
>>8650647theres one artist i follow who literally draws one character. id like to make a lora but haha
>>8650876ValueError: too many values to unpack (expected 2)
:-P
>>8650879What's your checkpoint file extension, ckpt?
>>8650879smells like you're giving the optimizer too many betas
>>8650880no i guess it doesn't like dashes in the name i had to rename it to 1
well the first steps look like shit so idk if that's just how it is or if my parameters are fucked, i gotta look further into it
>>8650881llm-kun we are not even talking about training now
file
md5: c6d9f7efe625e5dd2f3d1e289b982350
🔍
file
md5: 67ac4291631c5150430fc3e4c6efce59
🔍
Excuse the aco but catbox is down.
Well okay, at 4k steps there is definitely some stabilization happening, although it can get pretty grey with some artists. (I guess that's natural though, they're greyer in their actual images than you'd think, the Riz fits the color palette he actually uses.) (or is that a parameter issue???)
I'll bake a test on eps out of curiosity first, and then i guess i'll cut the lr to 20% and let it bake whole day tomorrow.
Actually, should i bake on not 1.0? I always thought 0.75/0.75s looked a bit better than 1.0 base at least for vpred, but I don't know how it'd impact compatibility and character knowledge.
>>8650887>>8650895sovl
Back fetish doesn't get enough content. Thank God for AI.
85682148
md5: 72f0fd3b8326c7a19804ec1a1d5c73c5
🔍
I love the generated pictures my LoRA make
i hope i don't actually bake a good stabilizer for eps
i dont want to retrain a gazillion style loras on eps lol
>>86509001.0 v-pred, looks are secondary to knowledge
else we'd all be using Tofu right now
>>8650900how large is the dataset you're training vpred on? also mind sharing a config? been meaning to do another finetune extract
>>8650914Is there actually any knowledge lost between the versions, that's my question. I doubt you can bake in much on 5% of the tune or whatever they possibly added in at.
>>8650916It's 8.5k of my best style images
Don't ask me for advice because it's literally the first tune test I've ever done but right now I'm using
>--resolution 1024 --train_batch_size 1 --max_token_length 225 --max_train_steps 10000 --v_parameterization --cache_latents --cache_latents_to_disk --sdpa --gradient_checkpointing --enable_bucket --min_bucket_reso 256 --max_bucket_reso 1024 --bucket_reso_steps 64 --bucket_no_upscale --full_bf16 --fp8_base --zero_terminal_snr --loss_type "l2" --train_text_encoder --output_dir "" --output_name "" --save_precision "fp16" --save_every_n_steps 1000 --save_state --save_model_as "safetensors" --optimizer_type "AdamW8bit" --learning_rate 0.000005 --optimizer_args "weight_decay=0.01" --lr_scheduler "cosine" --min_snr_gamma 1 --sample_every_n_steps 1000 --sample_prompts "" --sample_sampler "dpm_2" --dataset_config "" --pretrained_model_name_or_path --mixed_precision bf16essentially just the same as I've used with the style loras but lower lr, but like I said, for the full thing I'll definitely lower it much further down, and increase the steps.
With shuffled captions, 1 repeats, and captions filtered with the script and appropriate artist tags added.
>>8650900>Actually, should i bake on not 1.0? I always thought 0.75/0.75s looked a bit better than 1.0 base at least for vpred, but I don't know how it'd impact compatibility and character knowledge.0.75 exploded much more often
its actually the case for their whole noob vpred lineage
the further they baked vpred, the worse it looked but schizo contrast issues got less and less severe
"knowledge" difference between 0.75 and 1.0 felt very insignificant to me
>>8650921nta but idk how they managed to make vpred simultaneously look undercooked yet fried to shit
>>8650924I've said this before but considering even a very early epoch of a lora can fix some of its most apparent issues, I think they must've just had some error fucking up the weights at the end or something.
>>8650920>is there actually any knowledge lost between the versionsDo you think all that training from 0.75 or 0.6 to 1.0 was done on empty images? Just grab a few 150-image artists/characters and check. I don't have the old checkpoints anymore.
Ok nah, on a second look vpred-1.0 is the best, stylistically too
eps-0.75/77 are half decent but that's definitely not white background lmao
>>8650931damn resolution limit https://i.ibb co/JRd7kkqD/xyz-grid-0407-3453636.png
error
md5: cadbdcf16744f1ddbe3e4e7c67776c0a
🔍
why do nai extension inpaints give this error when I put the result back into the inpaint window? if I put the result in gimp, save with no edits, and then post that result into the window it works fine.
>>8650895cute back
for me it's nape
Considering 4k steps at 0.000005 was already somewhat working, I am thinking 0.0000025 for 25k should be about right? Thoughtserinos?
>>8650950you can use different LRs for TEs
>>8650955I guess but even for the loras I couldn't bother trying the fifty billion possible ratios, let alone here.
>>8650943sounds like some issue related to png stealth metadata
>>8650943>>8650963actually, no, it's just a bug where something gets passed as a string when it should be a bytes array
ive baked misamisa
mostly for the outfit consistency
consolidated it into amane misa, blonde hair, two side up, brown eyes, black dress, detached sleeves, black thighhighs, garter straps, black choker, cross necklace, cross earrings
https://files.catbox.moe/a6mct6.safetensors
>>8650705>surprisingly fast compared to lorasbecause it takes longer to apply an updated lora to the model compared to just finetuning it (when training a lora the gradients are less expensive though)
>>8650920>--fp8_basethis only works for loras by the way, it's ignored otherwise
i don't think i've done this pov before
or back shots overall
fun
>>8650814Well, that much is true
But I haven't decided yet whether it's worth having to deal with the shitty settings, shitty noise in early steps, instability, etc versus having shitty prompt adherence and slop seeping into my gens
And either no one gave me decent settings because you're all a bunch of autistic fucks, or there's genuinely no settings to fix these problems
Seems that doing a finetune for stability is the correct path but I have no idea how to do that
>>8651033Is that a somewhat roundabout way of asking how to gen on v-pred 1.0? These litterboxes are still alive
>>8648930But what it comes down to is using any kind of "snakeoil" extension that reduces CFG in the first few steps. CFG++ in my case, but others have used CFG Rescale, AYS and more. You could even just schedule it manually.
>>8651042>Is that a somewhat roundabout way of asking how to gen on v-pred 1.0?Kinda I suppose, I've made decent gens before but I don't understand the settings. I don't use comfy also so the litterboxes are useless
>reduces CFG in the first few stepsIs that to stabilize the weird noise in the beginning of every gen? Depending on the scheduler it can go up to 70% of the gen without making an image lol it's pretty funny but very inconvenient
>CFG++ in my case, but others have used CFG Rescale, AYS and moreI have no idea what any of that means, are these for comfy?
>You could even just schedule it manuallyThat seems incredibly useful, how can I do that?
>battle shounen
>good anime
>>8651048All available on forge/reforge as well.
CFG Rescale is a very old addon we used on the original v-pred furry model, EasyFluff. It also behaved the same way. I assume this turns CFG into some preset curve rather than constant.
CFG++ are new sampler variants that prefer very low CFG values, I stay around 0.8-1.5. No idea what it does under the hood but feels similar to rescale.
AYS is align_your_steps, a weird scheduler that fucks with the timesteps somehow.
Manual CFG scheduling can be done on comfy with either a custom node (forgot the name) or like
>>8648981. This one might be comfy exclusive, it's pretty niche anyway.
I'll leave the guesses as to why this helps to people who at least pretend to understand AI. But the original v-pred paper mentioned this early instability too, so I don't think it's anything the noob team did.
aiba-tsukiko https://files.catbox.moe/97jfg6.safetensors
amino kohaku https://files.catbox.moe/3665hi.safetensors
cawchan https://files.catbox.moe/mv4ejz.safetensors
chuujou kagetsu https://files.catbox.moe/bstyf5.safetensors
kannazuki okuto https://files.catbox.moe/wgzg45.safetensors
kingbawoon https://files.catbox.moe/vayg56.safetensors
kisaki yuu https://files.catbox.moe/8ru85i.safetensors
mio \(fastest lap\) https://files.catbox.moe/441atq.safetensors
i'm almost done with baking artists en masse i think
>>8651063>AYS is align_your_steps, a weird scheduler that fucks with the timesteps somehowit is incompatible with ztsnr by the way, if you use it you effectively discard all efforts that went into training vpred and you may as well just use eps
>>8651066oh, so that's why i had those results?
it's even worse
it blows up vpred even more
>>8651066Good to know, I suppose. I just remember people mentioning it as an alternative, around noob's v-pred shift.
>>8651063Hmmm, thanks for the info, I'll see what works for me
To be honest, despite all the shitty aspects, I think that vpred (particularly 291H, not so much the 1.0) is superior to every other model I've used, it's just overall very inconvenient
Despite very much preferring the prompt adherence (since I don't like any one style in particular), it's so good to be able to just prompt whatever I want, use whatever lora and settings I want and then inpaint whatever looks bad and be done with it
>>8651067>it blows up vpred even moreThat's weird considering it tends to turn everything into a grey soup
>>8651067it was derived for the old eps schedule with sigma_max=14.6 and if anything a vpred model with ays should perform like an eps model without noise offset or slightly worse due to compressed schedule during training
>it blows up vpred even morethis is weird, it should not do that
>>8651071>>8651073For me it really fucked with lighting, although I guess that's eps too.
karras is also not ztsnr-compatible
i wonder why localfags still haven't implemented shit easy fixes for vpred novelai implemented ages ago
>>8651065consider posting a collage of sample images along with these
at 50 pics I doubt anyone knows them by name
>>8651079Local died when panchito retired. We’ve been starving since.
>>8651079why bother? or alternatively, fix them yourself
@noob generate an image of asuka wearing this piece of "clothing" https://danbooru.donmai.us/posts/5678529
>>8651158>he actually did itI kneel.
>>8649798it's aight. i shitmixed it with noob vpred, colors got fucked a bit, but gained some character knowledge and styles got unfried a bit. i wish there was a proper way to extract a lora from rouwei
>>8649798I don't get it, are you supposed to use this for i2i or t2i?
>>8651173it's another ""base"" model so, you should use it for t2i but it's not that great (at least for me and my mixes)
>>8649798It's awful. They should stop forcing natural text down our throats and stick with danbooru tags only. Illustrious suffers from the same shit. You never get good results with any of the prompts. It's so inefficient and frustrating to use.
>>8651181The future is getting AI to write your prompts for you, gramps. Get with the times.
>>8650963>>8650964I disabled the stealth png extension and the errors stopped, but now the images generate with no alpha layer. I went to the nai api extension specific settings and saw that that had it's own stealth png setting that I assume was conflicted with the standalone extension. I set the NAI setting to stealth all images, and disabled the original stealth extension, and now everything's gravy.
>>8650943>>8651217nvm it just stopped giving me an error message in the UI. it still errors out if I try generating with the tainted image in the canvas. only corrupts gens like this about half the time.
>>8649910>edm2_loss_mmTried this on my fine-tune setup and it's much better than the grok one.
Well, it's just a single run on an extremely small dataset, but results look significantly better...
>>8651271can you make hera white boy
>>8651272mmm nyo? Enjoy my sloppy seconds, however.
https://www.pixiv.net/en/artworks/132285644
>post only AI slop
>somehow gets 10-20k views on every post
how does one attain this power? also can anyone tell the artist style for this?
>>8651282Isn't there a SEAnig mafia on pixiv that runs thousands of bots to boost the members' AIslop posts
>>8651282anyone actually on pixiv AI rankings are unironically part of money laundering schemes that boost their numbers via bots.
I'm not joking. This is an actual thing and it's fucking weird.
234
md5: 39480d84caecb82f0daf8ba221ae4d4f
🔍
see you on the other side
https://files.catbox.moe/9u9jei.toml
Can anyone help me improve this? I'm having trouble with small detail coherency.
>>8651312what's with that loss, lol?
>>8651357>caption_dropout_rateuse caption_tag_dropout_rate
also you can try nise and loss snakeoils like immiscible_noise and sangoi_loss_modifier
>>8651358>loss is weird after 30 stepsWAOW!?
>>8648979This is also why autotaggers trained on booru datasets will never improve anything beyond the mean tagged image. The only super autismo booru-style tagging I know of that could would rival e621 is the hydrus PTR (https://hydrusnetwork.github.io/hydrus/PTR.html). 54 million tagged images/videos is massive. Collecting the data would be the fun part though, since the database only stores file hash + tags.
>>8651379EVA isn't that bad though, the actual problem is that danbooru doesn't have tags to describe anything but the basics but then adds seventy different tags for a hand on the tits
>>8649563>>8649576>>8649606Finally some real vpred gens, shame with no metadata though.
tags are cope. natural language is cope. swarms of conceptual microdatablocks based on Platonic forms are the future. there will be 'atoms' which are colors, 'forms' which are abstract blueprints of ideals (hair-ness, clothes-ness), and 'derivatives' which are derived from those forms (miku's hair, 2b's outfit). there will exist a conceptual 3d layer that constructs the elements of the scene in the same way you can visualize an apple in your mind and spin it around. that 'scene of forms' is then rendered into an image.
>>8651390>EVA isn't that bad thoughIt still shits out huge breasts-large breasts and other related tags, even at p=0.55
>>8651396breast sizes are fucking bullshit with any model and are tagged like crap, heavily depending on artist's and interpretation of breast sizes
>>8651397That applies to other things too. I've seen EVA spitting out like 6 tags for a fucking aqua-colored bowtie, regularly.
My favorite was when the taggers would output say 10 tags for a basic hentai image but 40 for equivalent furryshit because e621 is that much more tagged
>>8651079No one can ever give me any reasons why why you shouldn't be using good old Euler A. People here make it seem like NoobVpred1.0 requires some crazy snakeoil to make it barely functional. Why are people so against Euler A + CFG 3? I don't get it, since I've never had issues with these settings on noob.
>>8651402>cfg 3i love floating clutter everywhere and mangled anatomy too
>>8651413did lil bro come here from 2022?
https://files.catbox.moe/k118cc.png
also a fucked up one because i liked it
https://files.catbox.moe/5krdd9.png
>>8651390My favorite is "gloved handjob" having a whole dedicated tag for itself.
>>8651390>>8651438When 99.9% of your users are prohibited from searching for more than 2 tags at a time, it's inevitable you eventually start getting explicit tags for implicit concepts.
>>8651362>immiscible_noise and sangoi_loss_modifierDo I need the dev fork for those?
Fucking asshole Lycoris dev changing weight decomposition defaults. 4 hours debugging why all my LoHAs were giving me noise output thinking it was a vpred issue. For anyone else who has trouble training LoHa/Con with Koyhass, editing the default to this fixed it. https://github.com/KohakuBlueleaf/LyCORIS/issues/246
>>8651392swarm metadata doesn't work well on with most metadata readers (like the op one), so I stopped sharing it cause people complained. anyway, if you want to replicate this, just take a gen you like, do a very rough lineart of your gen with whatever colours you like and add "black background" and "YOUR_COLOR theme". that's how I did these
>>8651444It's a finetune config, sry, should have specified that
>>8651443Are they broken even if the ui supports wd_on_output? Had to update reforge a few months back for that to work.
>>8651456No clue, I'm on the latest forge. I remember DoRa was broken when it came out around a year ago on both Forge and ReForge before it was patched.
>>8651448doesn't matter
https://github.com/67372a/sd-scripts
>local 3dpd faceswap is pretty much sota and very quick
>no desire to slop 3dpd
Truly sad.
The technology is real good though, WAY better than it was 2 years ago.
>>8651268Did some more tests and yeah, it straight gives me better results overall, but VRAM consumption is slightly higher for the same-size MLP.
>>86514673dpd AI still looks very uncanny to me
I would cum buckets to cosplayer gens if it actually looked close to real life but anything that isn't a 1girl standing portrait still looks awful
>>8651468Good to know it's not just me that prefers it. Saves me time re-testing since I forgot why I came to that conclusion in the first place.
>>8651467what's the name?
>>8651476facefusion
i fucked about with deepfacelab and it was way more effort
test
md5: 9f794f15f70389f0c26eae852b7baac8
🔍
hm
i doubt this will ever stabilize to the point of something like 102d but it could be an okay base for style loras, i'll try that in the full bake
>>8651474It also produces different loss weights, according to the training graphs, compared to the regular edm2 implementation for me. And these actually make sense.
>>8651481>picthese weights don't make any sense whatsoever
>>8651477is it a standalone app? no extension for reforge, right?
>>8651480Sorry, what are you doing anon? Lost the context and don't feel like doing puzzles today
>resume from state
>it just terminates once it reaches the first epoch save
uh okay nice :)))
>>8651516Seems cool, going by the images it looks like it improved 1.0 a lot while keeping the prompt adherence
>>8651284>>8651285This is different from the Indian one on twitter right?
>it just refuses to save if i loaded from state
uhm amazing so i guess it'll have to be 15k only
>>8651472>3dpd AI still looks very uncanny to meWell I hear chroma is pretty goo—
>but anything that isn't a 1girl standing portrait still looks awfulI mean yeah. I don't do much 3dpd anyway but I'm a huge fan of /e/ pantyshots, cleavage, and wedgies from behind so I check it out on occasion. Video gen is pretty good for this too and there's guys making thousands on youtube with it.
>>8651484Oh I need to remake this mix.
>very little change from first to latter steps
>102d still mogs
well, it was worth a try
but this again confirms to me that they must've just fucked up something near the end with noob if 1500 steps have the equal stabilizing power as 15000
>>8651575Where's the comparison with 102d
Anyone up to baking this guy?
https://www.pixiv.net/en/users/15047166
>>8651575this only confirms you're kinda retarded because you collapsed the output distribution
i'll collapse your hole your fucking faggot
>>8651564>wedgies from behindever gen anything quality with this?
Noob wasn't trained on danbooru pools was it?
How i can prompt this pose in nai 4.5?
>>8651621>cuck shit>nai userfigures
>>8651626Don't care about cuck shit. But pose, that's interesting to me. No idea how to prompt this pose. There is similar sex pose , but cant find decent prompt for just this alone
>>8651621Prompt the things you see and set positions.
>>8651631>>8651631Source? Target? That's new thing?
>>8651638https://docs.novelai.net/image/multiplecharacters.html
it's all here
>>8651631LOOKS LIVE NOVELAI WINS AGAIN!
>>8651282views on pixiv don't matter, just refresh the page and the view count will increase
New Anytest models
https://huggingface.co/2vXpSwA7/iroiro-lora/tree/main/test_controlnet2
>>8651661But anon, what about muh metrics? Muh engagement??
>>8651681Any info regarding the difference between A and B? Besides the old anytest works well enough on noob. If it gets any more strict I can just use promax union instead.
>>8651681I wonder if that model's creator feel like they would die the moment they write a readme
>>8651402Euler A adds a bit of noise at each steps of the gen. Euler A never reaches a "Truly" finished gen since the image is always shifting at each step of the gen. With enough steps, Euler A will end up mostly/only working on improving the noise it itself creates.
If you compare it to Euler, it doesn't add the extra noise. So at some point adding more steps barely or simply doesn't modify the image anymore. It reaches a "finished" result.
Personally, the thing I hate about Euler A, is the noise it adds might change the general direction of an image mid gen and your last step can sometime be mid shift.
I suggest doing your own tests. Pick a seed and do 1 gen for each step (Do a gen for 1 step, 2 steps, 3 steps...) and check how it works. You can also do the same for the other sampling methods and check which one you prefer based on these results.
>>8651825I did my own tests, that's not how ancestral samplers work.
What you're missing is doing the fist half of a refiner workflow, that is stopping the gen partway without changing the "target" step count. That's what actually changes the end result, now how many steps you do but how many you plan on doing total. Once you've settled on 20, the difference between 19 and 20 is no greater than when using Euler regular.
>>8651681holy crap which one do you download
anytest3 or 4? illustrious 2 A or B?
>>8651847ah of course
it didn't connect because he was looking at the cock and not his balls
>>8651851she's too distracted by something else
>>8651851his balls were missing. it's the main joke of dandadan in the first season
>>8651736>>8651849On twitter he just vaguely states A/B 3/4 have different datasets and approaches. It's likely just one of those try them and see things where it's hard to say which one is definitively better.
>>8651862Nice. Might actually try doing this in my story but sonnet is such a fucking prude.
>>8651873ths, while making this i thought about writing a vn or making a game, but i have to figure the images out first, like how to gen multiple character in one scene, and so on. new to this
>>8651879For me I'm following the fem protagonist around so it's typically just her with one or multiple men which is easy, although the last scene was her and her two roommates in a lineup that I decided not to gen. I like doing it this way because it helps me practice and learn tags. I picture some scene in my mind and start trying to recreate it, then load it up in sillytavern to fap to while the scene is ongoing.
not another f95 patreon scam vn pls
lol the new anytest actually trained on v2 of illu? THE broken model?
>>8651484FYI she essentially works out of the box on noob, this was mostly just for the outfit.
accidentally cool https://files.catbox.moe/pmnnky.png
>>8651824He initially wanted to operate only within the Japanese community, but his attitude has gradually changed. However, his basic stance hasn't really shifted that much. It's always the volunteers who end up making the manuals anyway.
>>8652002looks like something that crain1art would draw
Is there a way to schedule loras in reforge?
>find nice character design
>look at the art people made of her
>they're almost all 1girl portraits at a very specific angle, probably trying to copy the official art
>there's a FUCK TON of them
Christ.
>>86520961girl standing is an art on itself
>>8652097They're not even standing. Just portraits. I suppose it was likely because only a portrait source was available and people liked the design enough to make a bit of art just based on the one official image.
Are parenthases the only thing that needs escaping or is it also colons? Do exclamation marks, semicolons, question marks, and backslashs need escaping?
I saw someone mention a big release about a month ago that was supposed to be significant upgrade, was there a thing or was he just coping about some shit mix?
>>8652143just a shitpost. don't expect anything else from 4chin
>>8652143try again in 10 business days
>>8651591i thought this character was a dude from the thumbnails and i am now much less interested in baking
i gotta try rebaking my TES failbake lora
>>8652348wow what lora and model did you use
>>8652351No LoRa just Miqomix
I used novel-ai v2, with truck kun vibe transfer
>>8652362can give the link sir?
>>8652366>>8652366https://civitai.com/models/533446?modelVersionId=592934
local never recovered from pony
local never recovered from AOM
aom
md5: 577e5f30474511383ec9239bb9326ccc
🔍
>>8652377Cope schizo. It was a massive improvement.
>>8652377people did unironically get mindbroken by it
i still don't get how you can have a model that finally do most stuff and styles properly but choose to do butiful 2.9d
>>8652392butiful 2.5d indian woman models were a thing before aom too, shit like dreamshaper, toon* models etc
aom just was the thing that somehow infiltrated hentai scene
i'm still not sure why it was preferred over anythingv3 desu. maybe it performed better with nsfw due to gape60? prompting sex on 1.5 was still fucking awful though
less shitposting, more miqotes pls.
>>8652395nah it definitely did have much better coherence, 1.5 was just a total mess of a model
v3 was just a 1.5 sidegrade at best
>>8652400anythingv3 was melty but at least it didn't have the awful 2.5d-3d style that aom models and based64 had.
i actually remember liking anything4.5 despite it being memed on, was less 2.5d than pure aom2 but felt more stable than anything3. now that I remember this shit it's kinda funny that nai has literally same model names now (v3 and v4.5)
>>8652407yeah but these models could barely do 1girl standing without melties, stability was appreciated
and clearly people here have zero aesthetic tastes lol
>>8652416not like aom could do anything but 1girl either, prompting sex on 1.5 was more on proof of concept level rather than delivering something that actually looked good or fappable. easyfluff could do sex but it was the original furry sepiaslop model
>>8652424twas good enough to keep a thread alive for a year
>>8652426yes it was 90% 1girl standing
the tech existing was insane by itself and thread could survive on the novelty alone
nowadays you can gen actual nice-looking nsfw stuff without much effort but everyone simply got used to this
>>8652434i wish i could still like textgen as i did in 2019 because of the novelty
>>8652084one box for anon
https://files.catbox.moe/wvfia6.png
>>8652385omega turbo slop
>>8652328Neat, this actually worked this time.
>the elder scrolls, aela the huntress, facepaint, brown hair>the elder scrolls, deeja, argonian, scar>the elder scrolls, dunmer, grey skin, pointy ears, red eyes>the elder scrolls, flame atronach>the elder scrolls, serana, vampire, black hair, red shirt, black corset, black capeI tried including lydia and mjoll too but i think there's too few pics for them to work properly. As for dagoth and vivec, I posted them baked separately before.
https://files.catbox.moe/upzy4x.safetensors
>>8652564(Although admittedly this will probably be better baked separately, I mostly wanted to try it as a proof of concept.
>>8652564>flame atronachbased
>>8652385those both sucked, counterfeit v3 was the best.
>>8652594i'll actually rebake it with more stuff, baking separately was actually worse
could include the oblivion style atronach too but i think i'd just pollute the tag
>>8652599use a separate tag for the oblivion one?
>>8652564Multicharacter loras work now? Config?
>they always workedNo they didn't they'd just fuse the features together regardless of whether the characters were together in a picture or otherwise.
>>8652601i doubt \(oblivion\) would work that well desu, unless i did something completely different, i guess i could try though
>>8652605>>8633915
>>8652606That config is probably better for characters than style I guess.
>>8652609Whatever helps you sleep
>>8652564Just curious - what gpu do you use with these settings? Something with 8gb vram?
>>8652614>full_bf16>fp8_baseThen why?
>>8652615looks better
not even kidding, 1:1d it
>>8652617How many steps do you usually do?
>>8652606yeah it would probably have to be a completely distinct tag
>>8652619i don't work by steps because it's silly
It was a waste of time asking you anything unfortunately. The loras aren't even that good anyway.
that's it i'm hanging myself!!!
>>8652629He's right though. Here's a test I did recently. Notice the lack of activation tags or helper tags like traditional media.
Your config
https://files.catbox.moe/5x679h.png
Anon's config
https://files.catbox.moe/27hsa7.png
This one also went for far less steps.
>>8652714now post the config
All this talk of stabilizing v-pred 1.0, why did no one ever try converting eps 0.5 to v-pred instead?
>>8652728show us your eps 0.5 gens
>inb4 I don't use it BUT
>>8652728Give me at least 8xA100 and I'll try.
>>8652729Of course I don't use it, doesn't mean training the TE didn't fuck things up
>>8652730If it needs that much more compute then fair enough
>>8652728Isn't that exactly what noob did?
>>8652745well no, they converted eps 1.0
>>8652728why not just convert pony to vpred and end this madness?
>>8652748An anon already did though https://civitai.com/models/684052/v-pony
>>8652741>If it needs that much more compute then fair enough8xA100 is tiny compute today, literally babby tier
>>8652754brah why is the shinobou pic actually more 2d than anything out of noob
>>8652728little do people know, but there's actually a secret checkpoint which is exactly that...
noob is a failbake which is why people are still coping with shitmixes
>>8652765Indeed the only good base model is No*** AI.
>>8652767Yeah, Noob AI V-pred 1.0
>>8652714Wow, NTA but that actually does look better.
Post the lora rn or I'll shave your eyebrows
>>8652757...that much more compute (compared to whatever GPU(s) anons that have tried to stabilize v-pred 1.0 happen to have)
>>8652757This is why SaaS is the future btw
>>8652714Not bad, but I hate the AI gradients. They are in every gen that isn't flat color. When is it going to get painterly??
>>8652821>When is it going to get painterly??When you give us 16ch vae and a resolution above 1993 monitors
>>8652831Possible with 4ch though, notice transition between shades has a subtle texture? https://litter.catbox.moe/2bvxlv.jpg
It's everywhere, but I marked some with green arrows just in case...
Makes me feel very funny, but only very few loras / innate styles capture this.
This texture is actually even larger / less subtle at base resolution, makes the whole gen kinda melty, but it gets more balanced when upscaled, so low resolution is also not an issue.
>>8652714This doesn't even look like Bouguereau, just generic AI oil painting style
>>8652621of for cunt's sake
i was just testing it and realized all my atronach tagging files got all fucked up for no reason
welp time to rebake
>>8652850That's why I didn't even bother actually responding to it, I mean... come on... how do you slop 102d this hard?
>>8652717Yes yes in a moment. Here's another example.
His config
https://files.catbox.moe/22kyk8.png
Didn't do any keep tokens though.
Anon's config
https://files.catbox.moe/foc6r1.png
The anon who gave me this config was the one that originally baked this artist a while ago. This is his lora.
https://files.catbox.moe/txqn1z.png
https://files.catbox.moe/g26bov.toml
>>8652772Jokes on you, I'm already super saiyan 3. I think that lora is a failure because I baked it on the wrong model, which was actually my reason for posting. Loras baked on base illu seem to work "better" than loras baked on vpred when running on 102d/291H. That bouguereau lora in particular was baked on Illu 0.1 (not 1.0).
>just generic AI oil painting styleThat fact that it even looks like a painting is already an accomplishment. It won't look like the artist because he doesn't draw in 1 specific way based on the pics in the dataset, but it does get his bodies and the lighting correctly.
>>8652854Yours looks far worse so you seem to be seething. Instead of being mad why not put our heads together so we can figure this out eh? You love baking so this should be your interest and yet...
>>8652917Anon you posted (ai-generated:3.543) and (ai-generated:3.412)
There is nothing really to say
Guess you were never here for a serious discussion. My mistake. Come take the (You).
>>8652921you didn't give me the you you dunce
>>8652848What was the point of posting a jpg in catbox? It takes more effort to do this.
>>8652917>That fact that it even looks like am AI oil painting is already an accomplishmentnta but no, not really
>>8652924Original png without arrows:
>>8641954
Why is germie trolling here?
>>8652931You didn't answer the question.
post some scheisseporno to placate it
>>8652939That's just how I saved it when I added arrows, you can read into it all you like ¯\_(ツ)_/¯
>>8652917>just generic AI oil painting styleI don't think so, I think it looked good even before, you just needed some elbow grease. It looks better now.
Are you the one who made the Bouguereau lora to begin with or did you just get it to make a point?
file
md5: e93d29c4d834192c7c23cc44d8b549bb
🔍
>>8652945>Are you the one who made the Bouguereau lora to begin withNo, that would be the seething schizo that's mysteriously disappeared after being called out. I downloaded his lora to test configs and the strange stuff I've been seeing on bakes recently, namely that base illu works very well for some datasets but I don't know why.
>you just needed some elbow greaseHis lora doesn't look bad, I just find it strange that he even needed an activation tag or helper tags to get the style. Loras baked on 0.1 never do but loras baked on vpred are kinda weak. They need much more time to even get the style and still don't get it as well as 0.1 without helper tags. Ultimately I'm still trying to recreate picrel, but the faggot cutout the metadata for the lora.
nigga i am here what are you talking about
>>8652949Oh hey that's my image lmao
you ain't getting it nyow :3
>>8652949>namely that base illu works very well for some datasets but I don't know why.vpred probably doesn't have a single proper lora baking config to this day, that's most likely why.
Considering all the loss weighting stuff it requires (edm2) and other weird shit. If LAX couldn't get it quite right during their big bake, I imagine it's not simple to figure out for a single autist with a 3090/4090.
>>8652964it's momo_no_sukebe
>>8652975I wonder how much of stuff like o4 being technically impressive while still looking like shit is down to the people designing image gen models having 0 aesthetic sense.
>>8652975Would be a funny response, except nobody ITT achieved pic on the left so... it's more sad really.
>>8652983I don't think it's really that as much as them having to slop on a billion Bs of RLHF finetuning on the safest of artificial data so you don't accidentally gen a man topless
>>8652985dalle2 always had this chalky look. it just goes well with this example because it looks like it's intentional
>>8652975not the first time I see this but I always feel so much anger at the "significantly improves" sentence
if the prompt gave more way for chaos/model creativity I could bear it, but it explicitly says it should look like an oil painting
in what way is DALL-E 3 like an oil painting? it's typical digital art visuals
did those "researchers" never see an oil painting in their lives
>>8652961Yeah fair enough. Might as well go back to baking on illu then. The colors actually aren't that bad, especially on illu 1.0 instead of 0.1.
>>8652960I'm surprised you're this mad. What a show.
>>8653005>>>/h/hdgIf you don't want to discuss loras why not just go back? Don't you get tired of shitting up this place day after day? No one here likes dumb shitposters and we split to avoid your kind.
>>8652949>that would be the seething schizo that's mysteriously disappeared after being called outCalled out for what? I haven't been keeping up with the ups and comings of the drama lol
> I just find it strange that he even needed an activation tag or helper tags to get the styleIt's not unusual on civitai at least, I also remember distinctively needing to add "realistic" since it aids a lot in getting the style right.
>>8653012TLDR: he's mad I called him a slopping jeety so now I am a schizo
>>8651480will you be uploading the finetune when youre done?
>>8653023no comparison pic no proof. 102d affects some styles too much
>>8653025i mean... i guess you can wait for the proof of me never uploading it or posting with it then?
>>8652763Based cryptic anon.
>>8653018I fail to see how, at lest by the images he posted
It does look like it got Bouguereau a little better, not to say you didn't do a great job on it either. Illustrious/noobai models are notoriously shit at making oil painting and realistic gens
Thread bake? We crashing out.
>>8653028if you really want it it's https://huggingface.co/Laxhar/noobai-XL-Vpred-experiments/tree/main/sdxl_vpred_1c
>>8653032It's only page 11, there's still plenty of time.
>>8653036Time to merge back to /hdg/. At least we made it 10 threads, sisters.
>>8653032Give me some minutes
>>8653031I don't know what you mean
>>8653034>You need to agree to share your contact information to access this model
>>8653038I'm a tourist here but it feels like there's consistently a single tranny attempting to merge the thread back to /hdg/ everyday
its our civic duty, merge it back so hdg doesn't die
>>8653042>>You need to agree to share your contact information to access this modelhttps://civitai.com/models/833294?modelVersionId=962003
I think the thing anon linked also got uploaded to civit. Except there are two versions on huggingface and only one on civit. It's underbaked as far as i remember.
>>8653046you've linked shit ass old euge's test vpred bake he attempted on 7 4090s or shit, 1c is trained on the main cluster
>>8653046It's not only underbaked, it's broken, it was a proof of concept that euge funded out of pocket to convince LAX to bake a v-pred version. Also it was probably based on 0.25, not 0.5. I guess at least it's broken in a different way?
