Thread 8693972

Selfie Edition

Previous Thread: >>>>8689703

>LOCAL UI
reForge: https://github.com/Panchovix/stable-diffusion-webui-reForge
Comfy: https://github.com/comfyanonymous/ComfyUI

>RESOURCES
Lora Training: (NEW) https://rentry.org/yahbgtr | https://rentry.org/Copier_Method_Notes | https://mega.nz/folder/VxYFhAYb#FQZn8iz_SxWV3x1BBaJGbw | https://rentry.org/59xed3 | https://github.com/derrian-distro/LoRA_Easy_Training_Scripts | https://github.com/bmaltais/kohya_ss | https://github.com/Nerogar/OneTrainer
ControlNet: https://rentry.org/dummycontrolnet
Inpaint: https://rentry.org/comfyui_guide_1girl#up-to-date-comfyui-guide-for-1girl-and-beyond | https://rentry.org/fluffscaler-inpaint | https://files.catbox.moe/fbzsxb.jpg
4chanX Catbox/NAI prompt userscript: https://rentry.org/hdgcb

>FAQ
>What is model is everyone using nowdays?
Any of those
- NoobAI-XL (NAI-XL) (base model, BEST FOR LORA BAKING) | https://civitai.com/models/833294/noobai-xl-nai-xl
FOR DOING GENS:
- 102d custom| https://civitai.com/models/1201815?modelVersionId=1491533
- 291h | https://civitai.com/models/1301670/291h
- r3mix | https://civitai.com/models/1347947/r3mix-illu01nai-v-pred

>How do I do X
Given how the models work and are trained, you should always look for a corresponding tag in these places
- For anime -
https://danbooru.donmai.us/wiki_pages/tag_groups | https://danbooru.donmai.us/related_tag
See tags in action: https://tagexplorer.github.io/#/
- For furry-
https://e621.net/tags | https://e621.net/pools

>What I want to do doesn't have any tag or tag related to that
Time to train a lora (see the resources section)

>How do I add/remove/fix things
Inpaint (see the resources)

>I want to do X interaction and/or put Y amount of characters at the same time
You can try to prompt accordingly for it but you will end up using controlnet and inpaint (see the resources again)

OP Template/Logo: https://rentry.org/hgg-op/edit | https://files.catbox.moe/om5a99.png

shitmixes shouldn't be in the OP. Period.
Fuck off with that shit.

>>8693977
show your base vpred 1.0 gen with metadata

>>8693979
https://litter.catbox.moe/k69jgwk70kpas11z.png

>>8693982
sorry, but this is not exactly helpful:
upscale metadata (useless for base gen settings)
unidentifiable lora at 1.0 strength (might as well just be that slop stabilizer for all i know)

>>8694003
https://files.catbox.moe/hazb8u.png
>unidentifiable lora at 1.0 strength (might as well just be that slop stabilizer for all i know)
A single lora is a single lora. And all you need to know is that the dataset for it is from someone who is an inspiration to us all.

This one is really silly considering how the background is done, probably I should have dropped the sketchbook thing

There, 3 images

cant find good settings to inpaint a detailed pussy on r3mix...

>>8693972 (OP)
why do you keep rebaking this crap...

>>8694051
maybe don't use r3mix then
anyway once you have enough denoise for the change it's all art style

>>8694079
I got to a decent result I think, also did a lot of inpainting on the character and this is the final. I'm pretty satisfied with the results overall but I will definitely play around with more models, enough drawing Makima for now though.

>>8693972 (OP)
Why did you remove the ez-scripts fork, retard? nobody is using base anymore

>>8694184
He removed most of the useful resources and replaced them with a shitty FAQ

>>8694184
Because the shitty cumfy tutorial that recommends Illustrious 1.1 and WAI is more important.

>>8694184
because he's a /g/ vermin that doesn't know the actual resources and wanted to make an epic op with shitmixes in it, what else?

>>8694242
Thankfully it's just the split thread so no harm done.

Really wonder what anti-split schizos expect to happen if we returned to the main thread

at this point this thread gets more shitflinging by said gverminbot anyway so maybe don't come back to hdg?

>>8694250
Clearly late '23/early '24 /hdg/ would magically return. Everyone would just stop acting like assholes, parroting forced memes, spamming non /h/ gens, and freely share metadata.

>late23/early24 hdg
>good
What a newfaggot????

What dataset editors exist out there, anons? What do you use?

excuse me, groidie and birdschizo was actually peak

>>8694287
waifuset cause the dev is cute

>>8694287
Like tags? The webui extension for deduping tags, taggui for large datasets, and my scripts for automatic cleaning.

>>8694292
based
>>8694294
>taggui
Isn't that a little old? I think there were more recent UIs

>>8694300
i guess? it just works though

oh wait i just realized that taggui also has a tag dedup thing lol
can trash that shitty extension now

new VAE
https://www.reddit.com/r/StableDiffusion/comments/1msdho7/sdxl_anime_vae_update_they_are_the_same_picture/
https://huggingface.co/Anzhc/Anzhcs-VAEs/tree/main

>>8694324
Why is it so blurry?

>>8694324
ground truth where sir

>>8694329
here *spreads legs*

>>8694324
marginal improvement on colors and detail resolution, particularly things like pubes

Golden noise noob finetune will save local.

>>8694184
>Why did you remove the ez-scripts fork
Which one?

>>8694418
https://github.com/67372a/LoRA_Easy_Training_Scripts
and this
https://github.com/67372a/sd-scripts

>>8694420
None of those were on the "original" OP rentry but yeah whatever, I'll make space for them
>https://rentry.org/hgg-op/edit

Is there any way to freeze the te in ezscripts after some portion of training or is the only way to save states and manually keep track of lrs and training time

>>8694428
none that I know of, only save state then resume with different settings

>>8694324
do those still meltie when inpainting?

>>8694499
settings issue

>>8694324
well i tried it and it only made my oekaki meltier, not sharper

>>8694499
They shouldn't melt even with the default VAE if you aren't retarded

left is the "better" vae and yeah it just looks like someone ran animemeltie 6b over the lines

oekuckie bros...

There a guide for the tiled upscale meme?

>>8694025
>>8694027
Nice. Is this an img2img or did you prompt for the notebook and such?

>>8694263
>Clearly late '23/early '24 /hdg/ would magically return.
I guess you're a newfag like me. Still those were the days. Even when I showed up here in Nov '23 the oldfags claimed /hdg/ was already shitty and that earlier in the year or even 2022 was far far better. It would be nice if we could just get the '24 era back at least.

>>8694324
they're the same picture

>people unironically nostalgic for a graveyard era general
wowie

>>8694526
oh wtf i tried it
do people genuinely use that garbage?

>>8694526
The old guide was debunked and shitters never explain but I can post my settings.
>NoobXLControlnet_eps tile model
>tile resample preprocessor
>control strength 1
>start step 0; end step 0.8
>balanced
gogogogo (at 1.5x resolution). End step functions as a "fixer" as in the lower the end step the more radically everything in the image can change (and give you bad anatomy). The higher your end step, the less it changes regardless of your denoise. You'll need to play around with end step a lot, especially with loras.

>>8694559
denoise? also, straight img2img? half of these jeety guides used ultimatesd for it

>>8694526
I use
>1 denoise, 0.35 strength, 0.75 end
For the first pass
>0.85 denoise, 0.85 strength, 0.85 end
For the second pass, changing around things a little depending on the style. The first pass is rather aggressive, but it's generally beneficial in fixing the base gen's small mistakes, you can increase the CN strength and end step if your base gen is good already.

>>8694538
Pure prompt

>>8694559
That sounds more like upscaling with tile controlnet, not tiled upscaling.

>>8694563
I use 0.4 but use whatever you want. At 0.8 end step I haven't noticed much improvement with 0.95 denoise, just a lot of time wasting.
>>8694568
I don't know what you're referring to then. Ultimate SD upscale is shit.

>>8694564
Also as an added note, I noticed that generally, lower CN strength values result in thinner linework and a higer level of noise in the pic that may get interpreted by the model as more detail or just left there to be an eyesore depending on the artists and loras.

>>8694565
box?

>>8694571
>https://files.catbox.moe/l1lfbr.png

yeaaaah nah i can't see any discernible difference with either guy's setups compared to straight img2img
obviously it's more stable with controlnet at silly denoise strenghts but it predictably doesn't upscale any sharper

>>8694578
>obviously it's more stable with controlnet
Yeah that's how I would describe it. You're in full control and can shift end step and denoise around depending on your needs but quality wise idk. For some artists it's better but it feels like placebo. Regular img2img tends to give a lot of hallucinations and other bullshit so I just prefer CN and adjust the numbers as needed. It's very helpful when you're up at like 3500x3500 resolution.

>>8694584
I kinda want to try baking some artist at 1256 or 1536 or whatever and see if it somehow improves upscaling

>>8694578
>it predictably doesn't upscale any sharper
you have to raise your resolution even more my friend

>>8694591
that doesn't change the vae bro
but nah i tried two different ones and it's literally the same line clarity

>>8694569
nta, just saying they're separate things. He might not know what he was asking about either.

Using tile controlnet while you upscale is almost the same as using xinsir union or anytest, only difference is accuracy and which parts the CN focuses on. Tiled upscaling is meant as vramlet cope and for extreme resolutions. Breaks the image up into small tiles and upscales them one by one. So you won't crash on out-of-memory or start swapping to disk even if you go like 8K res, it'll just take a long time.

>>8694590
just use the 1536 lora extract

>>8694526
1.5x base gen kohya deep shrink
2x upscale, w/ realplksr https://github.com/Phhofm/models
downscale to 5MP total, denoise last 7 of 30 steps w/ low strength promax controlnet, using base gen as control image
downscale that to whatever fits in a 4MP png for export to 4chan

>>8694593
>that doesn't change the vae bro
the entire point of upscaling to insane resolutions is to overcome the shitty vae

our helpful chatgptbro :>

>>8694597
>realplksr
qrd? been using animesharp for 2 years already, anything else was either too shitty or too slow (dat upscalers)

>>8694593
The VAE issue is just requiring high resolutions to produce proper details, and requiring high resolution images (or close-ups) in training to learn those details. The way upscaling/adetailer/inpainting fixes eyes and other small details is by increasing resolution. That's all it takes, at that point the VAE doesn't limit you in any way.

rmao

>>8694526
use xinsir(from what little I've tested of euge's version it seems slightly weaker and you want the strongest controlnet so you can use weaker settings), .35~.45 weight(adjusting up when you need things to stick better), .45~ end step(this is long enough that it'll, largely, set the actual image but short enough that it'll be less constrained when filling in details). Denoise .65-.8, higher is preferable but again, if you're having issues getting it to correctly interpret things you can lower it to force it through.
ultimately though this is all something you have to actually play around with and use a lot to get a feel for since all the ranges can be varied for different reasons and different purposes.

>my favorite style from latest bake is 7 images
waow
>>8694596
I tried that, doesn't really look that good either.
I'll give the bake a try, obviously won't be good for base genning but it might do some weird shit to help upscaling, it'll take like 10 minutes anyway

>>8694595
>Breaks the image up into small tiles and upscales them one by one
Isn't this what multidiffusion does?

>>8694601
realplksr is much sharper than animesharp, much faster than DAT, i generally prefer the variants that have minimal noise handling but even the "real" one is fairly gentle on stomping detail

>>8694610
>much sharper than animesharp
that isnt necessarily good
I'll give it a try though thanks

>>8694575
Oh I didn't know crosshatching was different from hatching (texture). Ok thank you.

>>8694608
Yes, multidiffusion is one tool for tiled upscaling. Mixture of diffusers and ultimate SD upscale are older ones.

>>8694613
no significant ringing from what i can tell, the minimal noise handling versions are also better on avoiding banding
and the difference in clarity on final gen is very significant coming from animesharp

>>8694617
Which model in that long list?

>>8694324
it does sharper lines but also is giving me an insane banding effect (which i don't like)

>>8694610
"real" ones indeed seem to be really sharp and don't look too bad from the first glance, but "layernorm_gan" doesn't feel sharper than 4xanimesharp to me
i'll find a use for the real ones though

thanks again to the anon that recommended that dataset editor last thread. Idk why Tagger is so fucking slow but this one is super fast.

ok no the 1256 lora didn't help one bit
damn you vae

>>8694644
What dataset editor?

>>8694704
he's probably referring to >>8690197
i also added some other uis i managed to find into this list https://rentry.org/yahbgtr#editing, haven't checked every one though yet

kohaku just released another future SOTA model waiting for a bottomless angel to finetune
https://huggingface.co/KBlueLeaf/HDM-xut-340M-anime

>>8694736
how gud is the vae

>>8694736
tbdesu that actually seems like it could be really interesting.
>>8694738
https://huggingface.co/KBlueLeaf/EQ-SDXL-VAE
>it's an SDXL VAE
DOA

>>8694741
kek my shitposting wins again

>>8694741
>340M
>interesting
might as well hope that your toaster burns the image of a hot anime girl in your breakfast by accident

>>8694736
>1024px
>4chinz vae
just put me down already bros

>>8694741
>it's an SDXL VAE
it's not compatible with sdxl itself doe
that said that model is tiny, you prob won't get anything useful out of it

>>8694736
>webp dataset again
but why?

>>8694746
it's interesting because lighter-weight models ARE interesting. Not everything needs to be ultra-HD 4k BLU TEH REI tier resolutions.
>>8694748
still the SDXL VAE(finetuned).
also this seems even more dumb since smaller models would actually benefit more from a high channel VAE than larger models.

>>8694752
more parameters = better

>>8694753
yeah how'd that turn out for flux and chroma?

>>8694747
Well he specifically said he wanted poorfags to be able to use it. I've been telling you guys that models should be made high quality regardless of whether or not people can run it. If it takes 5mins to get your awesome 2000x2000 base res gen then count your blessings and download more vram. You'll enjoy kino one day. But if we keep chasing the "can I run it on my 6gb laptop??" retards then we're fucked.

>>8694751
Unless they compress it below 95 quality, what exactly is the problem with that? Storing all that shit in png is just a waste of space

>>8694752
>Not everything needs to be ultra-HD 4k BLU TEH REI tier resolutions
Yes they do, sweaty. If you want to gen in low res on that model you still can but if we do it my way, everyone benefits.

>>8694755
flux was sota when it came out and chroma was made by some retarded furry

>>8694755
>>8694762
did chroma flop?

>>8694755
poorfag animebros not being able to tune it is not a measure of success or not desu, it was the same shit of waiting 1.5 years for a half component sdxl tune (which was also derided at the time lol)
the plebbitors have been using flux for a while

>>8694758
you are training an image generation model, if you give it the same pattern for EVERY SINGLE IMAGE, i dont doubt its gonna start including it in every single of your generated images by default
more storage space for the data is gonna be maybe like 200-400$, and you are already spending thousands on training itself

>>8694758
Why gimp yourself unnecessarily is all. Just do it the right way for the peace of mind, you only need to succeed once.

>>8694755
Flux is literally local dalle-3 m8. It was only superseded by sora.

>>8694766
Yeah I understand that, but you can use lossless with webp and it would still take less space than png. And I'm not really sure if 95-99 quality compression "pattern" would get picked up by a 4ch vae at all. Noob's issue iirc was that they used like 90 quality, although I might be mistaken

>>8694762
no, people wanted to pretend it was when it wasn't. And then died down when people realized it was untuneable garbage.
also """"""sota"""""" means jack fucking shit in this space.
>>8694765
>n-no you're just poorfags! those are the problem!
oh fuck off you retard.
and I'd be interested in this shit because a model trained at 512x512 with a 16ch VAE on a pure chinese cartoon dataset with, ideally, updated *booru tags would ACTUALLY BE INTERESTING.
I couldn't give a fuck about "muh tekkknology!" dickwaggling bullshit. I want to actually make art with this shit.

>>8694770
yes i dont have anything against lossless webp, but there is literally no reason to use lossy webp other than saving a few bucks on storage, just why risk it
compressing jpeg images even further and compounding whatever artifacts they already had instead of actually trying to mitigate them is insanely retarded
the dataset kohaku is using is also 90, and i assume every trainer using webp shit is gonna use it as well

>>8694773
>no, people wanted to pretend it was when it wasn't.
It was the best generalist model for nearly a year and widely used for everything that wasn't porn.

>>8694773
We already have a 16ch VAE 512x512 model though, it's sdxl. If you downscale your 1024x1024 output images to 512x512, you'll get 16 channels per 8x8 square instead of 4.

>>8694776
too bad we're on a porn board then, huh.

>>8694770
>lossless with webp
just use jpeg xl, it's even better

>>8694773
>no, people wanted to pretend it was when it wasn't. And then died down when people realized it was untuneable garbage.
When it comes to prompt adherence it's as good or better than dalle-3, its main issue is awful styles without loras.

>>8694779
>>8694765

why are the retards from /g/ here?
FLUX is irrelevant. It has always been irrelevant. Even in chinese made AI advertisements they don't use FLUX. They use SD1.5 for that shit.
SD1.5 is unironically significantly more relevant to the world at large than FLUX has ever been.

>>8694770
>Noob's issue
i think the main problem was that the gpu servers were geographically located in china and they couldn't distribute the dataset super well (i think they were provided the cluster but not funds)

ugghghhh sdxl is so bad it can't do anime and was trained on 3dpd
(it's totally not deriding it because we are too retarded and poor to tune it lolll)
oh pony and noob? i looove sdxll!!!

>>8694786
Nobody cares about flux but it's simply a fact that models with more parameters are better if you can run them. Retards that want a low para model are invited to download jordach's magnum opus cascade 1B

>>8694786
You are very retarded. You're not discussing model quality but difficulty in usage. In fact, I know you're not stupid, you're just being dishonest by pretending you don't understand something so simple. Flux is too difficult to run and because of that (and because of the lack of competent local bakers) there was no finetune. If there was we would be eating good.

>>8694770
You are mistaked, that was Pony with the jpegs

>>8694765
>the plebbitors have been using flux for a while
and they have no taste, honestly.
The flux face is a real thing. Plastic skin. Just general lack of art style support too, not just anime, it can't do shit without LoRAs.
The only good thing Flux has is prompt adherence. But then you have to write a novella in your prompt to get anything good.
>dalle-3
I hate flux but it's actually better than dalle-3 and the only time you should compare something to dalle-3 is if you want to insult the model you're comparing it to. Oversaturated garbage.
Between that and 4o piss filter the image generation field is not, in fact, in constant evolution, it's devolving.

>>8694799
Flux should be compared to SD1.5 and SDXL, until we get an anime finetune the face skin etc are irrelevant because they'll change. Same with natural language prompts.

It's a good base when it comes to prompt comprehension, knowledge of everyday objects, sfw poses, etc. But wihout an anime finetune it's only usable for setting up /e/ scenes for img2img. A shame Chroma was only pretending to train on anime, it's apparently pretty good for photorealistic furries.

>>8694796
>Flux is too difficult to run and because of that (and because of the lack of competent local bakers) there was no finetune
If you think that's the reason for FLUX's lack of a finetune then I've got a bridge to sell you.
People wasted a lot of money trying to finetune it only to find out that it just cannot "learn" more than like, two things. You want to teach it a chinese cartoon art style? Cool! It can do that! Wait you want to teach it multiple DIFFERENT chinese cartoon art styles? Woah woah woah, hold on there buddy. That aint gonna fly. How about you teach it to draw a really cool ferrari instead of those other art styles, yeah? Wait. Stop. What do you mean you ALSO want to teach it how to do a Lotus Elise?
etc. Shit is hardstuck on what it can learn and how it will learn it.

today i will remind them

>>8694798
I remember there were some melties about noob using compressed webp dataset.
https://huggingface.co/Laxhar/noobai-XL-1.0
https://huggingface.co/datasets/NebulaeWis/e621-2024-webp-4Mpixel?not-for-all-audiences=true
quick glance through noob info and it seems like noob used 90 quality webp for the e621 part at least.
Pony's jpeg artifact issue was immediately obvious, gens were consistently noisy, and he did something like recompressing it twice or smth

>>8694805
SDXL had dozens of random normie finetunes fairly fast.
FLUX and chroma have both had nothing but failed attempts.
Also I'm not sure what your point is when at that time there was literally nothing you could do with SDXL that was relevant to this board.

jej

>>8694806
they used it for danbooru as well

>>8694809
I mean other than the boobs, he's not wrong. Turns out slightly more parameters is all you need.

>>8694808
there were couple of very annoying sai/cumrag shills pushing for the garbage sdxl shittunes, posts shitting on sdxl are likely a response to those shills.

>>8694809
In the end, people did use SDXL as bigger 1.5. Nobody cares about refiner and 2 TE bullshit that were part of sdxl

After testing the finetuning meme for a while I came to the conclusion that lora training is superior as long as you're constrained to <=24gb vram. The quality loss you get from shit like 8bit optimizers, full bf16 or stochastic rounding is just too severe.

>>8694815
>as long as you're constrained to <=24gb vram
What if you aren't?

>>8694808
it took quite a long time to start to get usable tunes for sdxl, i think ponyv6 at the time of release being considered the second coming of christ is pretty telling

>>8694814
>Nobody cares about refiner and 2 TE bullshit that were part of sdxl
Oh yeah, whatever happened to that shit? Is anyone still using that?

>>8694817
you could probably make some very nice lora extracts if you rented compute

>>8694804
because it's the distilled version

>>8694821
cue 50 different snake-oil salesmen popping up claiming they are totally undistilling it and it'll work this time I swear!
>>8694818
oh should also probably note that when SDXL came out a lot of people were still kicking with 10-series GPU's. By the time PDXL came out people had finally migrated to 30 series since it actually offered a tangible upgrade in cost/performance.
Meanwhile it's 2025 and brick and mortar stores are still stocking 12GB 3060's.

Can you guys post some gens between your model rambling so it doesn't look like I am the only one doing that, don't get me wrong, keep doing it just attach some gens from time to time
thanks

There won't be any breakthrough in image gen until compute gets cheap enough that a sufficently dedicated hobbyist can train/finetune a model without cutting corners.
You need to throw shit at the wall until something sticks to make any real progress besides scoring high on hermetic KPI dashboards, and at the current prices that's simply not possible for anyone but corporations which can't possibly give less of a shit about image gen, let alone about anime image gen.

>>8694832
Why would I post pictures of my wife for your to leer at?

>>8694130
different style but very nice I like it

>>8694834
which is why a smaller model is INTERESTING. It doesn't need to be something super crazy amazing displace everything sort of thing but at least people will actually be able to do something with it and that makes it INTERESTING. I think the approach here is dumb and bad since it's still an SDXL VAE but resolution is honestly a corner I'm "okay" with cutting, especially if you can use the param drop to instead utilize a heavier VAE.
I went back and trained some SD1.5 loras a month ago for shiggles(since naiv2 was open sourced) and man, I REALLY fucking miss how fast that training was.
>>8694832
why do you want to get me banned

>>8694832
we've been through this billion times already
people either don't gen extremely narrow "/h/-related" pics (penis in hag vagina, anything else is off-topic) or don't bother with doing "effort" gens that often, so the thread is just filled with text for the most part
rule trolls won, so those are the threads you get

>>8694836
this
posting gens is virtual cuckolding

>>8694846
banned for what lil bud

>>8694848
I am not expecting complete effort /h/ approved gens, anything goes really as long as it's not actually /g/ shit or something like that

>>8694815
full bf16 is supposed to increase quality isn't it
anyway 1:1 on loras at least gave me identical to better results on 8bit and full bf16

>>8694846
>use the param drop to instead utilize a heavier VAE.
That's not how it works. You need more params to store all the additional details the vae pic up during training

>>8694832
>>8694852
T-Take my peace offering...

>>8694861
/h/?

>>8694861
Oh uhm, it needs some work...! but thanks for posting

>>8694865
Bro... I didn't know pussy juice was /e/
>>8694867
I-I did work on it...

>>8694868
>I-I did work on it...
I, hmm, recommend zooming in and out a few times next time! gambare anonie!

>>8694849
If you post your wife, you're a cuck.
if you post another girl you're cheating on your wife
Only by not posting images you can truly love her

>>8694873
based vanillachad

>>8694873
I don't post my wife doe, I post yours.

>used 3060 12GB's are still selling for $200+ USD on ebay
this is hell

>>8694881
My wife is "skinny, slim legs, small breasts" fyi.

>>8694855
>full bf16 is supposed to increase quality isn't it
no lol, mixed is always preferable unless you're some sheik and can train in fp32

>>8694832
here you go sir

>>8694912
Is this magion?

>prompt for anon's wife
>only comes up as blonde hair blue eyes
intriguing

>>8694912
very based

>>8694916
it's godiva ghoul, (suenari \(peace\):0.6), (itowo \(itowc\):0.8), hatching \(texture\), with this lora at 0.4 https://civitai.com/models/1127584?modelVersionId=1528491

>>8694855
>full bf16 is supposed to increase quality isn't it
not really. usually you are doing mixed precision training (sdxl itself was trained this way) which means that the model is stored in fp16/bf16 and the gradients are fp32. full bf16 by itself converges noticeably slower due to precision loss in gradients..

>>8694815
what did you try doing?

>>8694924
Training a style. tried adafactor, lion and paged adamw8bit with a bunch of different settings. results weren't terrible and finetuning was always able to pick up some parts of the style that i never got with loras, but the overall result still felt kinda lacking in comparison.

>>8694917
How is that small breasts though?¿ Also hourglass unerotic.

can someone give me a higher definition of this pic?
A bit more mature girl, with small boobs and a large ass please.
Thanks, anons

>>8694932
>but the overall result still felt kinda lacking in comparison.
that's very vague, can you elaborate more on that?

>>8694943
adding another just to contriboot to the thread, but wrong pose

>>8694943
>>>/trash/slop has requests and billion sloppers hungry to fulfill them.

>>8694944
Anatomy sucked, backgrounds were incoherent, trouble with small details in general

>>8694937
because I wanted to actually post it to the thread without being banned
https://files.catbox.moe/fa5ric.png

>>8694949
just from the way you describe it, to me it seems like it's more of a dataset issue which is somehow less noticeable when training a lora (likely because lora training has much less expressive power)
>finetuning was always able to pick up some parts of the style that i never got with loras
>trouble with small details in general
you are borderline contradicting yourself there.

>>8694947
Tried but no luck? Maybe I'll follow your advice and try again anyway

>one of the stronger oemeltie styles just doesn't train in
hm

>>8694958
>you are borderline contradicting yourself there.
finetuning captured the vibes better or perhaps the sublte meltiness just helped with the more painterly aspects of the style.
>seems like it's more of a dataset issue
I tried different datasets and even included some face/hand crops but i didn't make a difference. I blame it on lacking precision

>>8694970
Uuoohh contrast

>>8694970
I fucking hate when I put "dark skin, interracial" in negs and bro comes out grey or yellow or purple or whatever the fuck except white. (I realize you probably prompted it this way)

>>8694976
No, he's just built like that. I don't really care about skin color for the dude to be honest, I kinda cancel it out when looking at a picture especially if it's POV.

>>8694972
>sublte meltiness
well, if you put it that way...i think i can relate. have you tried turning off/reducing layer strength when applying a lora? would be funny if this meltiness is somehow related to adaptiong to a tiny melty vae dataset better or accidentally frying some layers.

>>8694987
No, but I guess I can try that tomorrow

>>8694992
please post about it if you find something interesting.
actually that webp discussion before made me think...what would happen if you also compress your dataset using 90% webp? could it be noob didn't learn the tiniest details well due to the entire dataset being compressed?

>>8694921
Thank you!

>people actually schizoing out about "compression patterns"
i've been trying to bake in severely heavy jpeg noise into some of my loras and that didn't do anything

>>8695018
Ponyv6 (actual pony, not autismmix) reproduced jpg noise very well.

>>8694943
here
>>8694976
>he cares more about the cock than the girl with come on her face
kek
>>8695018
This. Even when I bake eencya it smoothed out the pixels.

>>8695021
What the. I guess that was the wrong one.

>>8695018
there's a good chance it's being filtered out by your negatives and/or banked somewhere in the model that recognizes it as something extremely specific and needs you to actually prompt for it.
illustrious models have a really, really bad habit of doing that to "traits" (whether you tag them in your dataset or not) like sketch, painting \(medium\), etc.

>>8695021
>>>he cares more about the cock than the girl with come on her face
Yeah that's because I don't just do "the cock".
https://files.catbox.moe/96olcd.png
Having a fucking n-word standing in the middle of a gen is revolting and not my style. Send them back to their own continent.

>>8695026
>negatives
shiggydiggy

>>8695027
American website so this is my continent, sir. (You) on the other hand clearly don't belong since you need to catbox your own pics, unlike me :^).

kiss no erotic...

>>8695035
I don't belong together with cotton pickers? I know. That's why I don't live in a place infested with subhumans. The internet is too global though, that's yet to be solved.

>>8695030
>using an overfit shitmix
>AND falling for dumb snakeoil
lmao

>>8695042
Exactly! So haven't you gone back to /b/? There's no coons there, only your kind do that shit.

this guy is trying so hard to fit in wtf
that's like 2013 edgelord talk

>>8695046
Same reason you haven't gone back to Africa.

erm that just happened

>>8695050
To fuck thick sluts that give it up for free? Huh?

ok now it's samefag

>>8695067
4chan is 1 person btw

>>8695068
have you seen the archive reposts in some of the boards? you're so right

>>8695066
Well. Good for you and that's your business.
I was talking about my _gens_ though before you or some other bro said "I shouldn't care". Well I care, I don't gen n-words. That's my business.

why is lilnigger afraid to say it on 4chinz?

>>8695074
Good to finally see some gens. Cheers.
>>8695077
Cuz he thinks I'm going to banish him back to /b/.

>>8695077
Having to wait 120 seconds is inconvenient.

>not a single comic in dataset
>artist doesn't even have any on danbooru
>gens a comic
AIEEEE

>>8695080
Don't use close-up

lilbro thinks he's a mind reader

>>8695090
>>8695092
perchance?

Does anyone know what happened to Sankaku?
Why is it always showing me the mobile site now?

>>8695095
i couldn't even load it a few days back so i think they're fucking with the site

>>8695085
I'm right.

>>8695105
Nigger.
Okay send me back to /b/.

>>8694324
...it looks worse? And he even has the gall to ask for money.

>>8695109
many such cases

But when I (ME) post Marnie...

>>8695118
I liked your gen so much I wanted to do my own. Amazing Marnie. Sorry you have to wait 3 days every time this reddit sub goes retarded.

>>8695120
>following the rules is reddit
You need to go back.

>>8695120
To be fair I was quite susprised, I think she looked old enough and nothing like a loli at all but I guess the style didn't help my argument that much

>>8695123
She was fine, the state of this website is not.

Inpaint detailing practice

>>8695124
what model?

>>8695150
102d

need more nigger gens here

>>8695194
Will you start screaming BIBIICIII too?

*ahem* BIBICIII

nose

painterly NEXT

I've been fucking around with the new-ish video gen in comfyui and despite working well this morning it's totally shitting the bed now (ex: dipping outside of allocated vram and slowing to a crawl for seemingly no reason when it was fine earlier). Anyone know what might be causing this?

Does anyone have the metadata for SparklingSenpai's gens?

geg

>>8695236
$25

>thread split to try and get away from shitposters
>but because this thread was a lot more active earlier the shitposters are just here now and it's indistinguishable from standard hdg garbage
waow
who could have possibly predicted this would happen

retard

>Man I like how sharp this old gen looks, what were my settings for base gen on that time
>the same sampler and scheduler combo that I am currently using but with 6 of CFG
Oh...

placebo

>>8695270
Why is everything but her eye blurry?

>>8695270
based orc god

>>8695018
oh well lets just compress our dataset twice with 90 jpeg then, surely nothing bad could happen

>>8695271
Crop factor to low on the inpaint, please understand.

dumb faggot

me?

Hello, can I gen a smarter, more detailed model?
I'm kinda tired of getting dolphins every time I prompt for dolphin shorts. And eyes being melted when genning something bigger than an upper body. And the model not being able to understand characters with asymmetric features, even while using LoRAs. And not being able to gen text. And... Well it's a long list. But you get it.

>>8695324
better learn to inpaint bwo, none of those are problems at all

>prompt 1girl
>CLIP now pulls from every tag containing the letters g,i,r, and l
FUCKING CLIP

>>8695327
What if I don't have a top tier GPU? What if I just wanna hit the generate button while jerking off?

>>8695333
haha well NovelAI only 25 a month my friend!

>>8695333
Then why are you complaining? lmao, are you one of those that want 1-click everything?

did someone say $25?

only 25 a month for SOTA anime and hentai model? sign me up!

>prompt doggystyle
>gives me a girl wearing dog themed clothes
FUCKING CLIP

Yeah nah, I'm going to sleep now, you guys deal with him

why does he say this so often?

>>8695333
>What if I don't have a top tier GPU?
You don't need a top tier GPU, any Nvidia GPU with at least 6GB VRAM is technically enough
>What if I just wanna hit the generate button while jerking off
Then you should either lower your expectations for what a GPU can hallucinate out of random noise, or wait some years until the tech gets there (or doesn't who knows).

>>8695357
>Then you should either lower your expectations for what a GPU can hallucinate out of random noise
but why can nai gen me some porn with one button while I jerk off and give me one gen every 6 secs but my 8gb vram gpu can't? that doesn't sounds fair

Is there any sort of inpaiting tutorial regarding NovelAI 4.5?

buy a gpu

>>8695399
What sort of "tutorial" do you need? there's literally nothing easier than this shit, no 500 toggles, no mask blur or other shit you have to care about, don't really have to change the prompt even if you are doing are inpaint, there's literally nothing to write about.

>>8695426
My question mostly regarding glitched fingers. Just mask them and inpaint? That's it?

>>8694976
You're using noob, right? Just prompt for the dick color you want. Negs were always a cope, and they somehow work even even worse on illustrious stuff than they used to on Pony or 1.5.

>sd1.5
>stacking loras
>spend hours adjusting weights to get right output.
>gens keep varying massively
>change "huge ass" to "massive ass"
>everything is fixed.
Wtf am i doing wrong.

What's with this sudden splurge of shitposting?

i noticed.

>>8695430
That's what I usually do. I didn't really have much issue with hands on 4.5 although I guess that depends on the artist tags.

>>8695436
i don't think anons should be allowed to post gens in this thread that don't have "galaxy penis" somewhere in the prompt

>>8695443
>Wtf am i doing wrong.
Using 1.5

>>8695445
>thread gets more active because there aren't shitposters to ruin discussion
>thread activity attracts shitposters
>shitposters frive regular posters away
>shitposters leave for more active threads
>repeat from step 1

>>8695462
God damn it.
I downloaded so many lora's for 1.5.
I'm just going to stick to Illust.

>>8695021
>>8695023
Man, thanks but as I was saving your previous gen you deleted it. The second is a bit different (6 toes, creampie). I literally was ok with the previous, that was cleaner, like this >>8695150
that sadly has no shown feet. But thanks anyway!

>>8695464
sounds reasonable but i don't see how this
>thread activity attracts shitposters
would work, why would they shitpost in a different thread?
also what the fuck is this
https://rentry.org/lora_train

>>8693972 (OP)
any online image gens just for doing anime-style backgrounds? That don't look like weird ass (tables with 5+ legs for example)

>>8695445
I'm new here, what amounts to shitposting itt?

>>8695456
kek
just showing off that you can do whatever you want with the colors, it doesn't have to be a tag

>>8695481
Try Flux on Civitai. Takes a lot of currency though cause it's a big model.

>>8695469
If your machine absolutely cannot handle SDXL, you should use EasyFluff with the HLL lora instead of older 1.5 checkpoints. It's v-pred, it's got enough sex knowledge for the basic stuff, like 700 artists and some mainstream anime characters and vtubers.

>>8695483
your post, because it attempts me to cite (and thus support) shitposters, disguising yourself as a newfag or an idiot or a returning oldfag to ask stupid questions, not having any respect for other people and posting slop >8695470 without even acknowledging someone might not like that, pointless shitting on other anons and trying to act arrogant, framing loli as "pedo", posting western-style images that should be on /aco/, "naiggers must die" and "it's just $25" kind of interaction, meaningless short posts without a reply or posts using gen alpha zoomer discord slang or any attempts to turn the discussion into a chat by any other means, etc

>>8695500
I probably can handle xl.
Its just i saw so many lora's for 1.5 i thought it was the way. Or i guess i should say I have a lora i wanted to use and it was for 1.5 so i've just been using that.

>>8695511
Thank you for the explanation.

>>8695517
Wrong file.

>>8695516
The reason there are so many loras for 1.5 is that you can train one in 5 minutes and that you need a lora for pretty much anything.

>>8695523
wow Did not know that.
Yea these gens weren't taking long at all to make even at 512 hirez x2 so i kinda started to think maybe i should go back to Ill.
Is ill good for semi realism ?
Maybe I'm retarded but trying to do short stack is so annoying and im running local.

what did he bite into bwo?

>>8695270
Never enough fischl

i mean... 25 dollars is a pretty fair price all things considered

less and less are saying this.

>>8695537
That advice was aimed at producing anime. You might be in the wrong thread, try >>8944432 or >>8925465 depending on the degree of realism. They generally use models based on illustrious or noobAI but pushed towards a more realistic look.

>>8695583
>>>/aco/8944432 >>>/aco/8925465

>>8695303
>ff14 autism
Nice. I made a lora of my character but I still haven't plapped her yet.

>>8695333
Unironically that was the point of novelai. If you can't run local for free* then you rent their gpus and use their model. If you wish to do local then you need to put a little more effort in that set up wildcards and generate, but not *that* much effort. Most noob models aren't too bad with things like prone bone, pencil skirt, or dolphin shirts like they were back on pony/autism. If you get one of those hallucinations just smudge them or paint over them in gimp and inpaint.

>>8695511
>framing loli as "pedo"
I love how you snuck this one in there, meanwhile my point was to tell you that loli belongs on /b/. You are still incredibly retarded for failing to comprehend such simple rules but alas, this is your elaborate shitpost after all. Gotta sprinkle in a bit of truth to hide your lie.

>>8695488
Can do glowing and other shit too I like to do ghost dicks plapping girls

>>8695470
The issue was that she wasn't old enough.

>>8695481
Flux by far.
https://files.catbox.moe/8ld8wp.png

>>8695572
Because most people can run noob, but if you can't it's better than doing 1.5.

>>8695644
i didn't say loli belongs here you fuckwit
>>framing loli as "pedo"
>retarded shitposter got triggered
checks out so far

>>8695646
wow thanks for the correction. saved. I've also got something from /trash/, in case I share with you since it's hentai still

>>8695644
Also forgot to add, loli is any girl under 25 and not obese. Since there is often misunderstanding.

Behold, the creative power of the SOTA anime model unleashed.

>>8695669
Why would you want skinny bitches? Are you a starving jeet or something? First world countries don't have food shortages and aren't scared of full-bodied women. Grow up rakesh

>>8695676
What I want is for me to decide. But bending the "loli" rule every which way because somebody posts what (((you))) don't want is how you got to the current state of this shithole of a thread. (but that's what you wanted in the first place)

>>8695678
It doesn't matter what I want or what you want. Moderators get to decide what loli and what isn't, and they decided that your post is, so they removed it for breaking the rules. It's rather simple, really.

>>8695680
No argument here. You got to the center of the issue splendidly:)

>>8695669
>loli is any girl under 25 and skinny
Might as well put up those consent forms and age of charas with each post so we know they're consenting legal adults

Bros should stop bending the offtopic rule and getting me banned for bugrape and tentacles, those were always hentai

>>8695699
At this point non-futa /d/ thread is less disgusting than this thread anyway and also has better moderation too. Might as well use that.

>>8695701
>second post i see is quadruple amputee medusa-stone fucking
amazing thread desu

>>8695703
I only said it's less disgusting than this thread, I never said it's perfect.

>>8695701
That really depends on your fetish tolerance, but it is comfier for sure. And not much slower at this point.

>>8695703
onahole transformation actually, he just used grey rubber in that gen

>half the posts are 832x1216 1.5 bullshit with watermarks everywhere
WAOOOOOW

how come it's so dead in these threads now? Expected a bit of an influx since doing short videos local is becoming way more doable

>>8695711
Thread was killed on purpose for the crime of being the best AI thread on 4chan.

>>8695710
the watermarks are an issue but I really never got why people here can be such sticklers about resolution, it's ai gen bullshit anyways so it's not like you're getting more detail so much as more room for artifacts and significantly increased gen time

>>8695711
Genning good shit takes skill and time, if you want a quick thread just go to one of the 500/300 slop threads for butiful 1mp sloppa

>>8695711
Almost all videogens (with real anime style) are bad or way too low resolution

>>8695711
/hdg/ is split in three threads right now, so obviously each only has a third of the activity. There's no going back though, fags will just keep rebaking it forever.

>>8695711
Speaking of video is catbox having issues with videos lately or something? I do basically 90% loli but when I try to throw webms up there they don't actually play.

I just remembered we don't have a new logo yet

>>8695713
That's true if you inpaint, or if you only gen close-ups. Otherwise you get shit like >>>/2993195 where everything's melted, her eyes, zippers, emblem on the cap, the gun, her fingers and most of the background. Upscaling is an easy fix for all of that.

>>8695721
I vote for it to feature black penis and a fat hag - the match made in heaven and the two most relevant topics for /h/.

>>8695713
That's wrong though. Doing tile CN upscaling or even genning at higher resolutions with something like kohya's hrfix yields much higher detail than genning at SDXL's standard resolution.

>>8695713
1024x1024 is 128x128 in latent space. If you are okay with that level of detail, check your eyes. (both yours and in the gens made at 1mp)

>>8695723
>>8695727
>some random retard says resolution doesn't matter
>something only someone who has never used imagegen would say
>so you decide to take it at face value and think it's genuine
i got a bridge to sell you

>>8695730
So you are saying all the people who post 1mp gens never used imagegen? That makes no sense, how are they posting gens then?

>>8695723
>>8695727
arguably the resolution of the final image doesn't matter. He may be downscaling the final result, or using a style that doesn't need fine details, or inpainting manually

>>8695731
i know you're intentionally obtuse but lmao at this strawman

link to some of those gens, I want to see them but I don't want to go there

>>>/d/11353427
answer my question

>>8695733
What I'm saying is your rose-tinted world where someone who says resolution doesn't matter is being retarded on purpose is wishful thinking.

>>8695735
>full nelson with stomach bulge
maybe /d/fags are not that bad you guys

>>8695736
aren't you a fruitcake

>>8695723
I just inpaint when I'm trying for something particularly solid so I guess that's why I feel that way, in my experience things tend to look worse upon upscaling but it might be something about how I've tried doing it somehow

832x1216 can be okay enough if you're not a tard genning on 1.5 and upscaling to that. Which these "people" are.

>>8695740
Link to your good-looking 1mp gens, bro.

>>8695740
Yeah, id depends. Think we'd have to talk about specific examples.

>>8695740
>in my experience things tend to look worse upon upscaling
that's exactly why you inpaint AFTER upscaling broski

>>8695743
nta but you can always downscale afterwards. Most 2K pics won't fit the filesize limit anyway, nor fit on my 1080p screen at full size.

>>8695742
yeah 1.5 is pretty dogshit, I honestly don't know how I used it in retrospect with how jank it always looks

>>8695751
These downscaled gens are obviously not the subject of this conversation.
>nor fit on my 1080p screen at full size.
4chanx dropdown toolbar -> Image Expansion -> fit width + fit height.

>>8695756
oh it was the other guy, my bad
I mean for locally saved gens, there's no point saving them larger than I can view without zooming out.

>>8695757
What if one day you buy a new monitor?

>>8695758
Let's just say I'm not thinking that far ahead with my gens.

Wait what, why are we on page 5 already?

>>8695784
>>8695554
slip'n slidin'

>>8695752
>I honestly don't know how I used it in retrospect with how jank it always looks
The mere fact of this tech existing was absolutely insane, even if it was just 1girl standing for the most part. It was hard to feel the limit of the tech when you didn't know shit and it was evolving every week, felt like you actually could do anything with the correct approach. I feel like it was Dalle-3 that disillusioned a lot of people in that regard desu.

>>8695752
for me it's the tinkertrooning, same as now
i still don't think these models are acceptable in quality

>>8695436
This just gives mismatching penis most of the time.

>>8695846
mmm potato chips

>>8695858
Multiple even

any way to prompt for push up? or tops that are otherwise holding up the breasts more than usual

>>8695879
just add "skindentation" bwo

>>8695879
use tags like perky breasts or bursting breasts or throw sagging breasts in the negatives

Sooooo. How do I prompt this pose on local?
https://files.catbox.moe/9ecb2n.png
Saved from /b/, has meta.

>>8695896
controlnet :^)

>>8695896
straddling, thighjob, from behind

>>8695896
girl on top,sitting on lap,looking back,pov,thigh sex,yokozuwari

>>8695896
https://files.catbox.moe/stnerz.jpg
1girl, loli, thigh sex, looking back at viewer, completely nude, leaning back, small breasts, lying on person, pov, 1boy, big penis,

>>8695903
tits too big

>>8695903
girl too small

>>8695903
this looks like it's got furries in it

>>8695896
Just get lucky

>rebaked the fugtrup with booru tags instead of tagger
>1000x better
fucks conventional tag wisdom fr fr

>>8695633
I used a friends lewd pic on img2img then did a bunch of inpainting to practice and I did learn some stuff while at it

>>8695928
stale b8

bait who nigga
myself and the bot's three alter egos?

>fugtrup
miss me with that tranny homo shit

>>8695959
yes, but also it's AI. Forcing artists to make shit they never would make irl is the main use case.

explain again what's wrong with being a tranny homo?

why are so many people posting hentai with children

>>8695972
they are children so for them it's just hentai

i made a script
https://rentry.org/on6tfm7e
first i grokked a script to query all danbooru artists with at least 150 entries
then i grokked a userscript to put N of them in at random into the clipboard, replacing underscores and escaping parens
you can set N, record a keyboard shortcut, and move the box around. default @match is novelai image and localhost:8188

for example

>>8696009
Shame the most interesting artists are below 150 pics

>>8695972
Truuueeee, we should have a conversation about that!
If teens are a no-no then why teens with big booba get a pass? Doesn't this promote inequality and sets unrealistic beauty standards for women?
They should be equally bannable.

>>8696010
you could add to replace the list with the complement, all artists under 150 entries
https://rentry.org/gciso7hw
this is my script for querying danbooru, maybe replacing
>'search[post_count]': f'>={min_posts}',
with
>'search[post_count]': f'<{min_posts}',
will do the trick

>>8696009
that's... a wildcard?

>>8696018
idk, i havent verified whether or not the artist fetch script does what it's supposed to
my guess is that line sets a filter lambda, therefore replace >= with < to filter for less than instead of greater/equal

>>8696009
shame inbuilt artists are ass

Are cute Britbong anons caged and have their censoring extensions installed? :3
You shouldn't be on 4chan, but gotta be safe if you are here.

Chat, do I wanna know?

>>8696022
>artist:nyl2
no way that does anything

>>8696031
4 pics on e621, unless they trained on webm thumbnails too
I'm guessing most NAIfags do not XY every token with limited gens.

at least make her anime bwo

>>8696009
example usage in my genning
now that i have this big list of danbooru artists, there's probably a better way to think about what i want to grind out and automate it, but this works perfectly for when i want a dice roll here and now

>>8696053
and oh, that thing i want is probably wildcards

why are mass sloppers like this

need zoomie haircut lora
need white lesbian haircut lora

how does one gen slop like this

>>8696133
stable diffusion

>>8696133
wai v140, big pp, pov, 2grils licky big coom

>>8695666
>Satan
>digits check out
kek the memes write themselves. If loli is not pedo then solo male is not /y/.

>>8696143
Wrinkly 110yo hands. Insides of her anus aren't in better shape either... Funny skin effect though.

>>8696143
ew

>>8696143
what controlnet does to mfers

>>8695798
>I feel like it was Dalle-3 that disillusioned a lot of people in that regard desu.
Wait what do you mean? That it made people realize how shit 1.5 was or it made people feel like local could never catch up? But yeah I feel like a lot of people here are jaded. I still marvel at the ability to see some hot obscure anime girl or gacha girl and just make a lora of her and create infinite porn. Even something as basic as pov standing doggy was a thing so rare in h-games and VNs that I'm glad I can just do it myself whenever I want. And then you add in textgen? Man idk why anyone is complaining desu.

for any schizos who are interested i've uploaded a bulk bucketing + dataset augmentation script for images with transparent backgrounds as well as script to fill transparency with fixed colors and caption them appropriately
https://mega.nz/folder/VxYFhAYb#FQZn8iz_SxWV3x1BBaJGbw
the goal is basically to streamline working with koikatsu and similar programs since this makes it not matter where the subject is located in your screenshots or what resolution or aspect ratio you're working with. i'm not sure if the resolution selection ratio is as good as it could be, that's something that i should probably test that out to see how close to the sun you can fly before you start getting wonky proportions

>>8695961
>2.27mb jpg
Did you just slap some film grain at the end or some shit?

>>8696166
I do 98 quality jpg without chroma subsampling when png doesn't fit in 4MB.

>>8696013
The real answer is that mods is gods and jannies are troons who will enforce the rules whenever they feel like it. Better to err on the side of caution.

>>8696143
Jokes aside, how did he do that skin texture? Guess it's the artist?

>>8696140
I wish it was that easy

>>8696136
indeed

>>8696170
Considering all the cellulite I'm 99% certain it's some 3DPD granny controlnetted.

>>8696170
There's paloma piquet artist that does skin glistening like that, but you'd probably need to train a lora to capture it. This looks like main artist was laserflip though.

kina crazy how good nai is honestly

noobai? yeah

>>8696180
true, novelai 4.5 (only 25 a month) is incredible!

wtf it's 2am in turkey go to bed

naijeets shilling their cowdung again

>>8696182
very true thats why active threads on other boards are all novelai 4.5 users while /h/ are all local luddites

>>8696180
is really good for complex cuckgens, I agree

reast obvioust samefaggus

>artist does little "when u see it" thing in every image
>even tagged they still show up in half the gens as little ghosts blended with the background
jej

proofs?

>>8696198
reminds me of dino back on NAIv3

>>8696198
>>8696206
>dino

how do you actually do lying besides viewer instead of viewer lying and the girl falling off the end of the bed?

>>8696212
You mean lying side-by-side with a male POV? I was trying to do something like that and looks like there aren't any booru images for this in the first place, so I doubt you'll be able to gen that (besides CN from 3dpd).

>>8696212
I just do lying on person these days. Cuddling helps too.

>>8696215
So I thought. I'm pretty sure I tried gathering a dataset for it before and getting like 3 pics kek
Would probably be somewhat forceable if someone made enough of a 3dpd/model dataset to bake over but eh I can't be bothered setting up modelling tools

>>8696220
why is her arm fusing into her ribcage
and why is her spine so absolutely fucked up

Awwww, dirty n-word feels cuddly. How cute.

>>8696222
Resolution. Next question.

1girl, disgust, leg lift, barefoot, looking at viewer

>>8696229
Almost have a gen like that

proofs?

>>8696230
I only have gens like that

>>8696232
bakaschizo?

>>8696140
closest I got tagging the shit out of it and running a lora with the artist style on hires fix

>adding negs actually seems pretty advantageous when using style loras compared to inbuilt artists
huh

>>8696236
The real redpill is that even this is lora dependent.

>>8696236
I said that months ago

>>8696238
Is anybody talking? Who's there?

>>8696239
hi~

>>8696236
Well okay it still softens the lines considerably on a closer look
Not that bad of a thing for some artists though IG.

>>8695667
>>8695668
>>8695903
box please

>>8696243
hehe nyo :3

>>8696243
is novelai 4.5 :D

>>8696234
further refined, I wonder if he just gachas until he hits a really good seed and what artist tags/loras does he use...

>>8696236
>decided to do a blind 1:1
>18:15
kek many such cases

decline?

my fault

I am raping Nobeta, sorry I can't post that here

what caused my chair to recline?

I found the missing piece of the puzzle that was eluding me,taking my gens to the high quality that I wanted. Using a hires fix workflow while pushing the resolution up as much as possible before the gens start breaking down has made the gens so much more insanely good looking its not even close to what it was before.

your lack of generacy of course

>>8696323
Yeah that's basically what I do as a 3 year old newfag. Plus gen in hires 1536x1024 from the start.

>>8696170
It's iuui + masamune shirow

>>8696325
>gen in hires 1536x1024 from the start
that's why her arm is fusing into her ribcage

>>8694130
box?

>>8696236
>negs
Negative prompts?

So here's how you evaluate training results automatically:
1. Split your dataset into 'train' and 'test', at least 5 images for test.
2. Train on 'train' part.
3. For each 'test' image take a prompt and generate at least 3 images.
4. Compute FID or SSIM or FID based on eva02 features between 'test' and generated images.
5. Pick median metrics.
Thoughts?

>>8696439
Just img2img of the picture above it, which is not mine. If that's what you wanted, here's the pre-inpaint https://litter.catbox.moe/3r416hasf3i52cwb.png

>>8696475
I would rather trust my own eyes, and don't see the point of automating something that only happens once every hour or so. But the theory is sound I supposed.

>>8696487
Well the entire point of automating this is to exclude any human factor and various small mistakes that go along with manual x/y plotting. And of course you can also do x/y plots and look at it yourself if you really want.
In fact, you can set up a blind test where you play the role of the evaluator by comparing generated images to 'test' images and make a leaderboard. But that's a second step of course.

>>8695154
still the best

>>8696475
That would just reward frying not generalizing you want

>>8696490
Depends, are you looking for the "objectively best" lora in terms of recreating the dataset, or "subjectively best" one in terms of looks? The style reproduction might not work out but it can still produce a cool style, better than the original even.

bots are stupid

I remember textbros once did a pretty big communal blind settings test with public results but lmao @ jmagebros 1) even having someone who can write a a website 2) participating

>>8696506
This is why you split the dataset into 'train' and 'test'. You don't train on test images. But yeah, in some cases if the data is too homogeneous then splitting it may not have a huge effect. Suggestions?
>>8696507
Well, to me this perspective is weird because I can easily distinguish raw gens from human artwork easily, so I don't think "subjectively best" is even applicable here.

>>8696516
It'd have to be something simple for one, SDXL still can't into actually replicating prompts
But still, I feel like that'd just be inadvertently become like trying to use loss for quality control, too divorced from life to matter.
And how would you even judge what epochs to test? Different datasets converge differently, and it all goes out of wack once you tweak settings.

>>8696519
wasn't this supposed to be akame

>check old gens
>try to port over old style I love into vpred
>can't hit it so try vpred loras same weights on the older model
>looks better
I swear vpred is a big psyop that I believe in like I can get stuff I like consistently but it always feels like its missing something, or the colors look too soft or something.

>>8696523
vpred is a meme because it was made for 3dpd not anime
so the ebin chiaroscuro doesn't really make sense in anime which essentially never used it

>>8696520
>like trying to use loss for quality control, too divorced from life to matter.
How is that so? SSIM and FID are computed on fully generated images in RGB space. SSIM probably isn't particularly good for content evaluation, but it's very cheap and you lose basically nothing by computing it. FID, on the other hand, uses features obtained from Inception V3. Which is not so useful for anime images, but it's possible to swap Inception V3 to using features from wd eva02 or camie taggers (camie was also trained to classify artists).
>And how would you even judge what epochs to test?
You don't judge, you test each one because all you lose is a little time.

>>8696521
Nah

>>8696524
>it's either grey sludge of eps or its fried chiaroscuro
Fuck off.

>>8696524
anime noise schedule when?

>>8696524
naiv3 is vpred

>>8696525
>How is that so?
Well, because it won't tell you how it prompts, how it gens different poses, if it has fry in that technically is how the artist does it but doesn't look good etc
The reality is once you have the baked loras with the different settings, the hard part is gone, visual result comparison is the easiest thing to do so trying to automate that out of all things seems a bit silly to me. (t. guy who did the 1:1 test)
>You don't judge, you test each one because all you lose is a little time.
Okay, but how much are you going to be training? Are you going to be adjusting them on the go depending on the settings or what? I'm just thinking logistics here. You can't very well compare epoch 11/1000 steps or whatever on 1e4 batch 1 and epoch 11/1000 steps on 1e5 batch 2
>>8696528
Kek it kinda do be like that
Drawn hentai is unironically pretty grey overall, or well, not as contrast and dynamic range fried as the typical AI is. I would say it's somewhat easier to try to force vpred to look more grey than it is eps to look more colorful for that happy median tho
>>8696530
NAI3 a meme

sigh
do i have to spam this thread too so they stop shilling paypig garbage

speaking of memes i gotta quickly retry that "just bake on illustrious for vpred it's so much better" meme

>>8696531
>Drawn hentai is unironically pretty grey overall, or well, not as contrast and dynamic range fried as the typical AI is. I would say it's somewhat easier to try to force vpred to look more grey than it is eps to look more colorful for that happy median tho
You're basing your whole perception of vpred by the poorly executed Noob conversion. A lot of shit went wrong with noob vpred, sadly, and its still salvageable despite that.
It doesn't have to be grey eps shit. Fuckton of anime arts are colorful, and the gens dont have to be chiaroscuro crap to utilize proper color range. EPS is nothing more than a technical limitation that is not in present in newer models. And despite your dismissal, what NAI3 did right was colors and looks. It wasn't great with prompt adherence but it looked pretty great, especially considering saas limitations.

>>8696533
I may be the originator of that meme. You do lose some accuracy but it cleans up burns and overfitting very nicely, so it's easier for shitters to bake that way. If you're less of a shitter it might be too much going all the way back to illustrious, my best lora was baked on v-pred 0.6.

Also, gotta say, if anyone actually wants to automate settings, it's gonna be much better to do a character as a test.
Styles even when well baked are limited by the res and vae and trying to discern whether a slight change in style looks proper divorced of the actual images is nigh impossible.
I love testing on Tezuka Rin because she is simple enough to not be broken on most gens but you still get the big indicator of her (lack of) arms and the possible sleeve bleed to compare.
>>8696539
Well one thing I can see already is that it didn't really converge in the same time the proper vpred bake did.

>try to switch to forge since reforge is old and slow as shit
>tile upscaling method on identical settings and models doesn't work anymore, causes distortion
alright

>>8696544
but at least regional prompter works again, right?

>>8696545
what's the point

>>8696548
four...five...six. Ah ha! Now I can call him out as a slopper

>>8696535
>the poorly executed Noob conversion
It's actually been executed exactly per ZTSNR paper, they also had this problem.

>>8696535
More like basing it on static CFG gens despite the v-pred paper itself recommending CFG Rescale. It was the same deal back on EasyFluff.

>try to switch to comfyui
>everything just works
>more than a year later
>everything still just works
alright

thanks for the bitcoins loser

I still don't get where the "lel just use rescale cfg" meme came from. It takes a single gen to know it doesn't really work?

>>8696552
But I HAVE to use Euler A Karras, anything else is snake oil!

>>8696559
I wrote a much simpler script that limited cfg to 2 or 3 for composition steps back when I used raw noob vpred and that actually worked well. Today I'm just using shitmixes and cfg++ and I don't care.

>>8696566
this is literally what cfg limiter does

>>8696551
did that paper describe adding cosplayer photos and frying text encoder

>>8696566
Isn't that what cfg++ does?

>eps, good
>eps not grey slop
kekarooooooooo

>>8696574
Are we still talking about contrast and color range?

>>8696531
>Well, because it won't tell you how it prompts, how it gens different poses,
Again, you DON'T train on test images. Test images (usually) have a different pose, unless the artist draws only one pose for some reason. The model doesn't see test images AT ALL, and you compare gens made with the model to test images. If you fry the shit out of the model, you'll get train images leaking on the test images/prompts.
>how it prompts
This is literally unquantifiable and exists only in your head.
>>8696574
>moving goalposts

>>8696579
I didn't say anything about test images though
That it'll do one thing doesn't mean it'll do another well, and prompting for test images or whatever won't tell you that, unlike manual prompting
I know for example that a fempov lora works on base but then just fries when using actual artist loras just because they weren't trained on that

>>8696579
>This is literally unquantifiable and exists only in your head.
also that's the point lmao
you're trying to quantify an unquantifiable quality of a thing that isn't made to replicate images exactly
i mean how does an test of "gen test image prompt" and try to compare it using algorithms to a said te(untrained) test image even work? it probably won't even give you the exact same pose or orientation as a test image unless you well, actually overfry it on said image
your approach doesn't make logical sense to the medium imho

>>8696533
Colors kept being fucked up for me. Now vpred bakes work great on eps it feels like too or better

>>8696580
>I didn't say anything about test images though
You seem to not understand the purpose of test images so there's nothing to discuss.
>it'll do another well
This is exactly why you test on test images which are not included in the train dataset.
>>8696582
>also that's the point lmao
What I meant by that is that you can't reliably tell if there's something wrong with the model either due to variety of factors including but not limited to picking a bad seed or you prompting the model incorrectly. Which results in you having coming to literally schizo conclusions about the model. There's no difference from automated testing in that regard.
>it probably won't even give you the exact same pose or orientation as a test image unless you well, actually overfry it on said image
This is why you are supposed to use FID and not PSNR or SSIM because it measures content, not pixels.

>>8696597
If FID can reliably quantify "stylessness" in different posez and resolutions between images then it truly is a godly algorithm. GL, I guess?

Besides, denying any useful data you get from automated evaluation outright because it doesn't cover all cases is just stupid and shouldn't be a reason you don't do it.
>>8696599
If you want to quantify styles, at the bare minimum you'll have to use a base model for FID which has some knowledge about styles. I imagine eva02 would perform badly in this situation because it was effectively trained to discard any style information. Camie tagger should be better at that https://huggingface.co/Camais03/camie-tagger

>you
isn't this supposed to be you doing it

>>8696613
just use novelai

>>8696613
I don't think ChatGPT can do that yet.

>single particular artist mix fries the shit out with my usual gen settings but it does fine with another sampler/shedule/CFG combo
Oh hell no... This is the second time something like this happen to me

>>8696641
means one of the artists is over or undertrained (probably over)

>>8696523
Yeah it's fucking weird man. Loras I train on vpred models work best on illu based models and gives me the improved colors.

only 25 dollars?

>>8696552
>recommending CFG Rescale
qrd?

>>8696533
>>8696539
>>8696541
>>8696595
Just post your findings and stop memeing. All my illu 01 bakes converged faster than vpred. In fact I only use vpred for characters because I can train it for very long without frying. The vpred style bakes would learn 80% and then fry after 10k steps.

>>8696641
dpm++2sa/sgm uniform is all you need

>>8696667
no it's not, I had to switch sgm uniform for beta since sgm was giving me bad colours

>>8696667
3m sde you mean with lots of steps

>>8696670
On 102d? I had the same problem with beta giving blurry bad colors which forced me to user euler/beta and I fucking hate that one too.
>>8696672
>3m sde
Hmm...
>lots of steps
How many steps?

>>8696673
>On 102d?
yeah, dpm++2sa itself is fine but sgm uniform was not giving me the results I wanted

>>8696677
What cfg and steps?

>>8696673
i use 50 but if you want less go for 36

>>8696665
>10k steps
nigger wtf are you doing
my 400 image datasets were 2k steps

>>8696683
It's called an exaggeration, anon. The point is that vpred doesn't fry for me, but it also didn't learn as much for whatever reason. Not like this conversation matters since you won't post your settings nor even your results.

>>8696679
5 of CFG and 30 steps

>exaggerating train steps for no reason
uh yeah i betcha?

>>8696630
oh yeah i guess it never said it will do it, it just said some random bullshit and "thoughts??"

>>8696681
>>8696685
From my quick test (1 picture) 2sa/beta makes things a bit sharper and improved brightness rather than saturation for colors which isn't bad. 3m sde/sgm is subjectively different. Either you like the way 2sa makes pics or doesn't, but 3m sde/beta was also a lot brighter than sgm. I might just do 2sa for normal and 3m for hires fix portions.

>>8696700
yeah 3m sde is nice at detailing

>>8696684
I'll agree with you vpred is harder for it to take. My settings worked fine for styles from eps to vored but characters I had to like double or triple my rate to finally get it to take

>>8696667
I need to just fuck around with latest reforge I guess so I can properly decide if your meme'ing or not missing all these samplers

>>8696663
https://arxiv.org/pdf/2305.08891#subsection.3.4
>We find that as the terminal SNR approaches zero,
classifier-free guidance becomes very sensitive and can
cause images to be overexposed.
After which they propose CFG Rescale as a solution. These days we also have CFG++, dynamic tresholding, cfg limiter, etc and even manual cfg scheduling. Point is to have very low CFG in the first few steps, then ramp it up to a normal value.

I can't believe the real answer was to use eps 0.5 all along

>>8696704
>with latest reforge
I heard someone say last time that 2sa was removed or something? 2sa has been there since the beginning so it's not "new" or a "meme."

>>8696706
i unironically would if someone shitbaked it properly

actually i should try replicating 102d myself
i now know how

>mixes
>eps
lmao

>>8696700
Hmm I don't know. They're both just as fuzzy but slightly brighter colors? I can just use cd tuner on the hires fix pass to improve colors. This will be a lot of work to test.

how do you get the girl's head to lay down on a surface? in this case a bed

Try the comfy versions of samplers if you haven't before, they're way better

>>8696735
head on pillow

>>8696737
then that's the pillow not the bed. I want her to lay completely flat

>>8696738
head on bed

>>8696736
Hey friend it seems that you were so excited to share your findings that you forgot to attach the X/Y plot proving your statement about image generation! Don't worry, it happens to the best of us and we are eagerly waiting for you to share it so everyone can see it for themselves ;)

>>8696741
What are you poor or something? Just do it yourself.

>>8696740
Wowie anon let's not get ahead of ourselves here.
That kind of thing... It would only work on NAI.

>>8696743
apparently, she's sleeping with her head straight up in the air, this shit is so fucking busted

>>8696742
Wow no need to be rude, but that's just how this kind of discussion works you know? You provide some evidence to corroborate your claim and then people are able to judge for themselves.
If you could just make unsubstantiated claims and then ask people to do the experiment themselves, it would be chaos! Also, it would open up the doors for trolls to spread purposefully malicious information, and that's not something we want in our community!

>>8696748
I am not discussing anything doe? Is a fact.

>>8696749
But anon you didn't provide 「facts」just 「opinions」, 「facts」require evidence to substantiate them, otherwise they are 「conjectures」 or「opinions」and you cannot plce the burden of proof on others. Hope that helps!

>>8696752
>and you cannot plce the burden of proof on others
Proof?

>>8696743
>>8696745
head down

>>8696745
>her head straight up in the air
Pic?

>>8696757
Kek.

>>8696743
>That kind of thing... It would only work on NAI.
It does!! People were shitting on T5 a lot but honestly it feels like the thing that carries 4.5 despite its very meh aesthetic

>>8696745
I can get it pretty reliably, but only with negpip (pillow:-1), else it will always show up under her head
just on stomach, on bed, head down

>>8696745
1.5kino...

>>8696772
What about from side, on back, profile?

>hey wtf base noob with rescale really does look way better
>realized the model didn't update and i was still on 102d
jej

>>8696772
I negpipped pillow and the whole bed disappeared, now 1girl is sleeping on wooden floor.

>>8696778
Some of you still deny it for some reason but that shitmix is legit good enough for most part, perfect for genning

>>8696772
same prompt and seed, just added your stuff https://litter.catbox.moe/iy0yxp0ychsl9v3u.png

neg head_rest too if you want her arms out of the way, or better place them somewhere else intentionally

>>8696778
yeah definitely doesn't make base usable out of the box
good for 102d though, definitely helps fix that vpred fry-in even when using low cfg, i'm keeping it on
>>8696780
it will always mog others just by the fact of being loras instead of cattower merged with easyfluff and wai 10 times over like the others

>>8696778
So, does base not look better with it?

>>8696788
>>8696784
i guess a bit regarding the contrast but it still doesn't make it more detailed or anything
that's with a style lora

>>8696776
>>8696779
What do you mean? Like this?
>>8696778
lol
lmao

>1.02d page only has the latest shittier version recipe
>civit not on wayback machine
>no real merge metadata
n-not like this..,

>>8696784
catbox??

>>8696794
1girl, sitting

>>8696795
I wanna see your rescale settings, obviously.

>>8696772
just negpip it for the first 30%, same effect but doesn't bork your whole gen most of the time

[(pillow:-1.0)::0.3]

>>8696796
the default out of the box

>>8696793
lol
lmao!
:skull:

>>8696784
Not a great comparison, they all look way too similar. A strong artist maybe, or heavy quality tags, or a lora?

Here's mine. Base, rescale, cfg++ and 102d custom

>>8696789
That's what people said about loras stabilizing noob, you don't need anything else once you're using one even at a low strength.

>>8696805
isn't it the same with yours
rescale on base only added hand melties everywhere

>>8696805
>102d custom is just an overpowering garbage style from all of the merged loras
I will never understand shitmix users

loraless it doesn't really change that much either on base
yeah sure 2% better but it's not 102d-usable tier to me

>>8696814
base looks like looking at a miniature at 300% zoom with how shitty and unfocused the lines are

>the reason people use shitmixes is because they can't bake their own shit
>and don't know how to find loras that they actually like
>so instead they just rely on whatever shitmix has the strongest built in (AI-generated:1.21) output they can find
everything was a mistake
maybe there hasn't been any image gen advancement because you heathens don't fucking deserve it

>>8696810
why is she wide eyed?

>>8696820
Good. Only you deserve to make good pics and those pics are for your eyes only. You've ascended.

>>8696040

Seeing her anime-fied looks weird to me is all

>shitmix bait
wtf have i taken a time machine to may?

>>8696814
>>8696817
In the end all that matters is how nice it looks not how faithful it is to the style. It's like that other anon talking about eyeballing epochs. Your goal is to pick out the one you like the best and sometimes it isn't the one that's 100% faithful.

>>8696808
at least the XY shows it doesn't do much compared to other options

>>8696825
Can you like not reply to me with these nonsensical non-sequiturs? Thanks!

>>8696828
Huh?

all that matters is that it's a placebo piss-water that only changes a few small details while retaining the AI generated quality we've all known for so long and grown to deeply respect and love

>>8696820
My style mixes looks almost 1:1 on base noob against 102d, the thing with base noob is that I need to roll way more than I should for a good base gen

>>8696821
it big

>>8696834
WHY IS MR T. WHITE?!

>>8696828
Ohhh... you really are the worst shitposter here. Even my shitposting is higher IQ than this kek.

>>8696835
ouch

>>8696834
zangief's knuckles are fucked

>>8696837
that's what communism does to a mf

>>8696841
Hmm, true, I didn't put too much care into it at that time

>>8696834
WHY DID HARRY DUBOIS GET A MOHAWK?!

god damn
whatever anon mentioned rescale i thank you

>>8696847
cheers

>>8696847
jesus talk about a conclusive blind test

>>8696849
what are you using

>>8696850
https://github.com/rainlizard/ImageBatchCompare

>>8696854
wtf that's great

>>8696856
quite
i'm pretty sure an anon made it too, i think he shilled it in hdg or something and that's where i picked it up

>>8696854
that's actually really cool but i also wanted to know about your rescale settings

>>8696860
>>8696798

>>8696864
What happened to Serena’s clothes?

>>8696868
Lost them

okay one bad thing about rescale is that it does quite exaggerate the background greyness i get with some character loras

>>8696878
not just with loras, lol

>>8696880
it ain't that bad without chara loras at least
definitely goes full on grey on white backgrounds with characters, i wonder why

>>8696882
nah dude i'm runing x/ys right now and rescale is consistently grey

>>8696884
me too and it's not that bad lol

oh it's when it's a chara lora with a style lora that it shits itself
either alone are okay
weird stuff

>>8696887
when i turned the cfg up with rescale on i got distortion that looked like when you got too many loras applied so i presume it operates in a similar way somehow

>>8696890
From what I can see the difference with rescale using one lora is literally nothing backgrounds wise, but when I put in two+ of any loras it just utterly breaks into grey meltie
How weird, I bet that's some kind of a bug.

What if I was already using the skip early cfg setting? Why bother with rescale?

>>8696475
Nigga just now thought of validation splitting. It's been part of ez-scripts for a year, lil' bro

>>8696909
yeah i already use guidance limiter (same thing pretty much) and it's nice

>>8696490
machines aren't qualified to judge artistic value

>>8696909
>skip early cfg
oh man that's fuckin weird
what value do you use?

>>8696915
0.2

>>8696916
real strong
it's the ignore negative prompt thing in settings, right?
>>8696912
gotta check it out
rescale is definitely bugged