>>8887702
How can she be so bored and so sweaty at the same time?

>>8887705
Because she's a lesbian and she's struggling with imagining it's a girl with a strapon

File: requests.png (3.43 MB, 1432x2144)

3.43 MB PNG

Request anchor :D :D :D :D :D

>>8887705
She's currently in Europe

>>8887708
this is some early ayakon tier shit
you really don't have to prompt for high heels and gloves when she wears them by default, they'll just end up on everyone else

>>8887705
it's like that meme where the guy is trying not to fart next to his crush. she forgot to poopoo before anal sex

File: 1723889280122835.png (917 KB, 768x1280)

917 KB PNG

>>8887699
removed the onions

File: prodiggyconstant.png (44 KB, 785x1047)

44 KB PNG

>>8887503
I'm gonna need peer review on this before I bake, it'll be 110 epochs, 40 images as I don't know how long it'll go.

Thanks for the daily help guys.

File: 1756087186024207.png (2.51 MB, 1824x1248)

2.51 MB PNG

>>8887702
>forgot to remove the turbo lora from the prompt after deeming it shit and reverting the rest of the settings
...

>>8887718
i assume this is the one with the turbo lora?

Realistic tags are actually kinda helpful with anatomy, background and composition, particularly if you're not going for a pure animu style
Maybe tag prompting will be useful

>>8887719
No, the VERY painterly one from earlier is. This is the intended look.

>>8887722
wow

Why is having low shift a bad thing
I've seen anons talk about it in these threads but in my personal tests it looks better than having higher values
Has anyone done a x/y/z plot on this?

>>8887716
For ProdigyPlusScheduleFree, delete all optimizer args except weight_decay https://github.com/LoganBooker/prodigy-plus-schedule-free/blob/main/prodigyplus/prodigy_plus_schedulefree.py
It's a schedule free optimizer so just set the epoch to something like 1000 and test every epoch until you get a good epoch then stop training

>>8887728
thanks, and if I were to do normal prodigy, it'll be like that setting I shown?

1k epoch is i'll assume 24 hours though.

67

>>8887729
You don't need to train for 1000 epochs. Train until you get a good epoch

The new turbo lora is ass

>turbo lora is ass
you do not say

>>8887734
i'll bake straight away up till 90 epochs like usual and see the findings, it's not like I can stop it and then gen something so I must see it to the end.

>>8887736
I agree
The only upside is that it makes inpainting easier somehow
I suppose it's a good thing to have in handy when you want to test composition really fast

>>8887734
nta but is there a convenient way to do that? save state, try out the lora then resume, and manually repeat as many times as it takes?

File: sav.png (20 KB, 1147x232)

20 KB PNG

>>8887741
I have it ticked in now hope it works.

>>8887741
Yes, saving state and then resuming is the way

>>8887742
iirc it takes a ton of disk space, I wouldn't save after every single epoch just pick a place to stop and save there

>euler_a: Softer, thinner lines. Can sometimes tend towards a 2.5D look. CFG can be pushed a bit higher than other samplers without burning the image.
Yet retards in here will pretend euler a is bad

>Can sometimes tend towards a 2.5D look.
blud. your own quote just explained why euler is SHIT

Euler A is not bad, it's ASS

>>8887746
a

>>8887746
What do you mean 2.5D is peak

>>8887746
I wonder what they meant with that part. It smooths out gradients, but you don't have gradients in the first place in actual 2D. Could it be referring to 3D stuff?

>>8887745
euler A is shit because it doesn't work with controlnets if you use them the way you're supposed to use them and it's shit for upscaling.
genning straight t2i at higher res it's "fine." not really great or preferable but it's "fine."

>>8887751
sounds like bullshit, upscaling works fine too
and how are you not supposed to use controlnets

>needing to use controlnets
literally just write nigga it works every time

>>8887754
>just write
I'm brown

>>8887751
imo depends on the style you're going for
I know people here hate anything that isn't anime coloring, flat color, (oekaki:2.3) but euler a 6 cfg works nicely with soft shading and some artist tags

Any recommendations for what LTX model to use with a 5080 and 32gb RAM?
Mostly just looking to do i2v

>>8887757
use the english dictionary csv for tag autocomplete
and search for stock images they usually have a description you can copy

>>8887760
LTX is abysmal dogshit for anime, it has absolutely zero idea what it's supposed to look like and will make it into uncanny 2.99D aberrations within just a few frames. Use Dasiwa (WAN 2.2 shitmix) with the general NSFW LoRA if needed. Layering LoRA on top of LoRA really fucks with the details so you should only use the bare minimum, which with an NSFW-capable shitmix is just that one LoRA for when you just aren't getting enough motion.

File: 1764827456508109.jpg (411 KB, 1792x1294)

411 KB JPG

>>8887759

>>8887760
the daisiwa one I tried just like this guy: >>8887763
It takes me about ~2-3 mins just to gen but I have a 4070 ti + 64gb ram though.

Also that guy's base workflow is bloated with unnecessity things tf you need voices for, etc. but at least it works.

What's up with the shift next to cfg when using anima. It's set to 3 by default, should I change that?

>>8887764
cool lizerd. box?

>>8887766
nobody knows what it does, like clip_skip back on SD
some low-level shit that's supposed to make the model behave "different"

>>8887769
clip skip isn't even relevant anymore but it's still treated like it is by some people.

>>8887753
>upscaling works fine too
if you understood how euler a worked you'd understand that it works in a technical sense but it's pretty bad for it.
>and how are you not supposed to use controlnets
euler a only really "works" for controlnets if you have them on for the full generation duration. which you shouldn't do. Ideally you have controlnet end at 40-60% step count but if you do that with euler A it'll basically just erase what you got from the controlnet and do its own dumb bullshit rather than properly building off of it.

>>8887772
It does nothing on architectures other than SD1.5 (and possibly old SD1?). But some UIs fill it into metadata, Civitai asks for it when uploading pics, etc.

>>8887772
iirc it's just cumragsoft that cares. sdxl needed it flagged to -2 or something to work properly? I think. but on any a1111 based UI, clip skip does absolutely nothing unless you're using sd1.5 based models.

>>8887778
Not SDXL, SD 1.5, more specifically the NAIv1 leak.

>>8887766
I don't know the proper explanation but in my experience it's basically where in the gen it distributes most of the noise
If you set it to the lowest, it will immediately gen most of the details and the artstyle very early on, and then it will slow down as it refines it further
If you set it to the max, there will be a lot of meandering noise early on and it will slowly build up the details while accelerating towards the last steps
You would think this is straight-forward, but not really. Setting it at low will make the artstyle and the details look better but can fuck up your anatomy and composition and vice-versa, plus this affects how you use prompt editing greatly

>>8887767
Which one

>>8887783
could you send some examples, maybe an xyz plot? I can make my own in a second but I'm running another experiment

>>8887786
i like both

>>8887783
Due to how distributing works, if you set shift to high things like
[@artist A::0.8], [@artist B:0.8] will still have a relevant effect in your gen, but setting it to low will only make 0~0.3 or so actually relevant

>>8887778
-2, zero, or nothing
Anything but 1 really, because that's when it tries to enable clip skip and it's not meant to be used with models other than SD1.5. Webui/Forge prevents you from being retarded and just ignores the setting no matter what you put it.

>>8887787
Sure, but it will take a while my hardware is kinda bad
Will be a good experiment since I'm using an autistic artist mix with prompt editing
>>8887788
https://files.catbox.moe/zyzp1u.png https://files.catbox.moe/65wj4g.png

>>8887782
does cumragsoft use - as a denotation or something and not as "negative?" because yeah nai leak/subsequent models used clip skip 2, but I mean that cumsoft breaks SDXL if it's set to anything other than -2 it would break, as a sort of function of the software itself.
but no, sd1.5 models used clip skip as a sort of "layer" function. There were some really funky fucked up furshit models that tried to contain specific features/datasets to individual clip skip values.
>>8887791
yeah that makes sense. And yeah, a1111 ui's completely disable it for anything not sd1.5, since it doesn't work/isn't intended there to begin with so there's no point in allowing anyone to enable or use it.

>>8887792
You and me both. My poor 3060 has been through so much.

>>8887774
That makes no sense. You've already pushed the gen elsewhere than the seed would have converged to, so when you cut off guidance both samplers continue operating as if you'd just started an img2img with partial denoise from that point.

>>8887795
I have a 3050 with 8gb vram lol, the average gen with 30 steps takes a min and 20 seconds or so
It is most unfortunate that a x/y/z plot doesn't accurately represent what goes on with shift since you can't visually see noise throughout the gen like I can since I have live preview on
https://files.catbox.moe/d73tb2.jpg
Check the artist mix to see the prompt editing

send nudes

>copy good anima gen prompt
>change one setting or tag slightly
>suddenly shitty gen
Is it just me or this shit finicky

>>8887805
remember noob base model?

Maybe I should start posting some of my slop gens here after all. Thread has so few image posts these days it might still be fun to see for some people around here.

>>8887702
slop

>>8887718
SOUL

>>8887815
do it faget

holy fucking shit people in this thread do not know what is clip skip anymore

File: ComfyUI_41814_.png (1.12 MB, 1024x896)

1.12 MB PNG

>>8887822
Is it that surprising? Any oldfag who's been here since pre-SDXL days you would be able to recognize from their posting or gen style, it's been years.

>>8887815
This is the only way forward. Though you can't revive the thread on your own, you'd have to get other people to genpost as well. Also should probably do it in /hdg/ and end this farcical split, it's gone on long enough.

>>8887797
>both samplers continue operating as if you'd just started an img2img with partial denoise from that point
and what do you think euler a does?
oh right it tosses more noise on top of it and you just lost everything you forced it to try and do(since that added noise allows it to try and correct itself back to what it WANTED to do), while regular euler will be forced to build off what it already got to.

>>8887830
It doesn't add additional noise on top of what was already denoised with the controlnet's help. Those steps are done.

>>8887831
>It doesn't add additional noise on top of what was already denoised with the controlnet's help
it does, in fact, actually do that.

>>8887833
This is a strangely common misunderstanding of what ancestral samplers do. At each step they remove some of the remaining noise and replace it with a new random distribution. The denoised potions of the image remain untouched. The practical result of this is they converge to a different result whenever you change the step count, because that changes the noise schedule it pulls this "random" distribution from.

You can check this yourself as well, using Comfy's Advanced Ksampler. Stop the gen with a couple steps remaining and decode the result, you will see the already denoised portions do not change as it advances through further steps.

Thoughts on the anima turbo and highres aesthetic boost loras? I usually stay away but this time around they don't seem all too bad.

>>8887716
SchedulefreeProdigy isin't as good as came/rex combo, this one repeats certain hair waves, higher epochs shrink the head while elongating/widening the body, including adding the hollow circular highlight I don't even have on the sameples unless I increase the lora weight as high as 1.5, came/rex will just either do one of the two things amost all the bakes, I'm reaching the goldilocks zone for came/rex anyways, but doesn't necessarily mean I won't test other optimizer/training combos.

File: shift.png (31 KB, 1000x600)

31 KB PNG

>>8887766
shift influences how quickly the model removes the noise per each sampling (or training) step, with lower shift (e.g. 1.5) you almost immediately skip to low-noise steps, meaning composition is going to be ass because you gave very little steps to it. with higher shift (e.g. 7.0) you spend more time in high-noise region.
also note that shift is the only "correct" way to adjust sampling schedule of flow matching models, "beta", "karras", "exponential", "ays" or whatever were not made for flow matching and when you use them there's a ton of math fuckery going on without a good reason.
>>8887769
clip skip is a hack introduced in novelai's first anime finetune of stable diffusion (they resumed from sd 1.4), basically it cuts some last layers of the text encoder before sending embeddings to the diffusion model. supposedly the model has richer embeddings to work on that way. sdxl adopted this trick for both of its dual clip text encoders from the get go iirc, and you don't need to set it manually.
>>8887745
>>8887830
the only thing euler a does differently from normal euler is that it removes the noise it initially added and then back puts the same amount of different noise. sometimes this may fuck up sampling with weird sampling tricks. maybe it may fuck something up if the model was trained with, for example, optimal transport (which tries to avoid training on very very bad noise in order to make the model rely on noise more) or some other noise energy bullshit. i'm highly surprised but this anon is 100% correct >>8887840

>>8887847
>back puts
puts back

The nudes?

>>>/e/

You fucker motherfucker
Once i get back to my home you'll see

bloody bastard sir I beat your families cow

gift card idea: chudai aunties

little brown bro thinks he's all entrepreneurial

the great thing about being indian is I only need to sell 2 AI images a month to make the average wage!

No lies spotted

Unlike SDXL, Anima uses the Qwen text coder, offering the significant advantage of being able to output images with the intended composition using natural language prompts (such as images with multiple people or composite scenes).

If WAI's image quality and Anima's prompt understanding are combined, it will become the standard model for the future.

little brown bro tried

>https://files.catbox.moe/cg2ucb.png

>https://x.com/hulixiaobai/status/2059306763741266196?s=20

>https://x.com/rushicrypto/status/2059360925787046174?s=20

>>8887893
yeah AI let's me use my coding skills to make art, a hobby I always liked but never had the time to get "good" at because I have to afford rent.

>your best was good enough
no it wasn't I grinded shitty drawing for years I was never going to be good. Although it does leave me annoyed at people I am far better at doing traditional art complaining about AI stealing their work because they don't understand the tech,

lul

>https://x.com/firedotalt

File: me methaporically in the front.png (3.7 MB, 1640x1808)

3.7 MB PNG

bros.. how do I deal with knowing this kimchi slut will never remotely admonish my little brown existence with such unbridled contempt while draining me into a prune with her grippers?

wut cfg best for anima

not too low
not too high

>https://civitai.red/models/2621858/realanima-x?modelVersionId=2980523
I found the Preview 3 version to be excellent for cg vidya sloots.

>>8887934
why not base1

>>8887915
bout tree fiddy, ironically

>>8887937
Haven't tried it yet but it's what I linked. Just sharing P3 was p good so hoping base1 is better.

(chibi:-1.8) in positives with negpip

File: original[1].png (1.07 MB, 1440x1072)

1.07 MB PNG

here's your local novelai reference, bro
https://www.reddit.com/r/StableDiffusion/comments/1totumo/anima_can_edit_images_and_this_is_possible_in_two/
https://civitai.red/models/2654416/anima-can-edit-images-or-testing-anima-edit-loras-and-ic-methods?modelVersionId=2980586

>>8887950
P-pancho.. *cough* forge neo support onegai.. *cough*

>>8887952
forge neo isn't panchos, you mongrel

>>8887953
Yes, nigger. I'm aware. But HochiMing02 isn't thread culture.

File: t202.jpg (3.36 MB, 1728x3024)

3.36 MB JPG

>>8887883
>>8887930
Give SDXL some rest bro...

cucks

ducks

bucks

>>8887958
anima sucks doe

yep yep yep many are saying this

anima desperately needs e621

File: 1769920818623633.png (774 KB, 832x1216)

774 KB PNG

>>8887950
So he just described the NAI inpaint trick from the pre-reference times and then highlighted two nigh unusable LoRAs extremely limited in what they can do.

>>8887971
I guess the main takeaway is just that the inpaint trick works with the Lllite inpaint controlnet.

>>8887972
I thought anons here established that shortly after the LLLite Inpaint model came out?

>>8887972
old news >>8882511

File: out.webm (2.66 MB, 1353x1920)

2.66 MB WEBM

>>8887971
>trick from the pre-reference times
The trick is still largely relevant and wasn't replaced by char reference. For example reference works like shit with inpainting model (they even have a warning for it) + anlas cost is gay.

File: catbox_4p8z48.png (3.28 MB, 1792x2304)

3.28 MB PNG

>>8887894
catbox? I really like the style

>>8888003
cry

File: 00074-3515761564.png (3.01 MB, 2400x960)

3.01 MB PNG

>>8887958
use case for not using SDXL?

>>8888020
anything besides 1girl standing. but we don't do that here

>>8888020
I want a robot gf so bad, bros...

File: 00026-1818123197.png (2.09 MB, 1920x1248)

2.09 MB PNG

>>8888024
>anything besides 1girl standing
any example?

>>8888026
nope sorry can't think of anything

File: 00003-2551090214.png (3.11 MB, 1664x2432)

3.11 MB PNG

>>8888027
based

>>8887958
ani said that anima is heavily shilled by comfyanon's matrix server. the hype is almost entirely artificial

File: t203c.png (2.98 MB, 1152x2016)

2.98 MB PNG

>>8888020
New = better.

>>8888037
can you get a new brain then pls

/asdg/ seems gone for good >>>/aco/9208012
Likely a follow-up to /r/ deletion which was due to deepfakes.

>>8888046
good riddance

>>8888048
Losing a containment thread hurts everyone else.

>>8888046
can't they just move to /s/ or /gif/ or /b/ even

>>8888046
I was going to say they missed one >>>/aco/9196124 but in the time I was making this post it got nuked. oops.
(also I legitimately do not actually know what the difference between that thread and asdg was at all)
>>8888050
only if the content it was containing isn't banned sitewide.
if it's banned sitewide, losing a containment thread doesn't matter - there's nothing to contain.

>>8888053
/asdg/ was local, mostly Chroma, Z-Image and SDXL merges. MMAI was about trying to trick various SAAS options into doing nudity.

yeah but good anima gens doko

>>8888055
sounds like a dumb and worthless endeavor.
also it seems that their futa thread is mostly realism so that's kind of... funny. wonder if that'll also get nuked.

>>8888057
>>8887958
>>8888037

:skull: :skull: :skull:

when will /h/ finally ban 2.5d bullshit? if /aco/ can remove realistic crap, so should /h/

p*interlybros, our response?

>>8888058
No big dead, it's been slowly dying for like a year anyway. I would put up a gachaslut cosplay once in a while.
Just feels like the site is getting shittier every other day, and this is another step.

if only there was another site without these issues which already has bunch of cool people from the thread postin :^)

>>8888066
isn't treechan dead? the link didn't work last time I tried

every alt-chan is always dead, come join our thread with 1 post a day

>>8888063
Why would they ban the only hentai that looks good? Nobody wants to see primitive doodles made by unskilled losers, much less masturbate to them.

>>8888070
shut the fuck up and consider suicide you faggot

Bery angery...

i love masturbating to doodles

>>8888046
Never went there, but I assume they started to generate unironic pedoshit.

>>8888086
wrong, 4chan is just getting rid of "deepfakes" and 3dpd ai in general

nice. very dead.

File: 00224-2953476747.png (2.21 MB, 1280x1536)

2.21 MB PNG

https://files.catbox.moe/3oa34h.png
https://files.catbox.moe/otqa35.png

weow

i miss fernus

Surely NAI v5 will fix the "precise reference" that is worse than the original character reference

File: 00242-3524458080.png (2.39 MB, 1280x1536)

2.39 MB PNG

>>8888107

File: 1755019600882578.png (731 KB, 749x1095)

731 KB PNG

>>8887985
I still use the inpainting when i want to make nudes

sir, where are the big heads, can't shitpost about that shit

File: 18-04-2026 17 37 28 - ani(...).png (1.15 MB, 1024x1536)

1.15 MB PNG

>>8887819
>>8887827
Alright, you've convinced me. I'll post some of my slop gens.

File: 1769327660022143.jpg (332 KB, 1792x2304)

332 KB JPG

>>8888107

>>8888113
welcome back, i've been extensively doing different loras lately but considering your more used to gloras, (to an extent 100% of my noob loras was based on your grand design since noob 0.5.), how do you combat with overfitting? the color, body is near on spot as it can be but the moment I do "from side, all fours, etc" it'll do exactly how the body is (i.e one hand lifted up, one hand under breast) despite not adding in the prompt, the one i'm doing is the came/rawr combo lately in the past few days never had that problem.

File: anima patterned texture.png (1.46 MB, 826x977)

1.46 MB PNG

>>8888113
I know everyone will have different scaling of images with 4chanx so I'll just screenshot mine.
Do you see the grid pattern when on a zoomed out version of the image?
This exclusively happens with anima. Basically when it generates any kind of grainy/noisy small texture it repeats in a patterned way and a crapton of images look like this zoomed out. Basically why I have to try my best to generate the cleanest shit possible with anima and add noise myself if I really want it.

>>8888125
on my screen only the fern one does that, your screenshot right now and the original, i'll assume the type of texture/grid stitch a real life canvas has.

>>8888127
Real life canvas will not have imperfections that perfectly repeat themselves every square cm. And it happens with non-painterly images too.

>>8888125
the grid thing is absolutely something I'm aware of/have known about, and I notice it when checking the images with windows photo viewer when it's not zoomed in on a 1080p monitor.
it seems like it's just an anima thing? It happens on "full background" outputs.
this also isn't dataset dependent. I've seen it happen on other datasets I've trained on. It also doesn't always seem to exist, it just happens sometimes when you output at larger resolutions. It's also possible it might be associated with one of the media tags? would need to test for it I guess.

>>8888131
Yeah it's not a dataset problem because for me it often happens without any loras. Can also happen with just a single artist with no quality tags and media tags at all. Maybe bigger resolution triggers this more but not sure.
It seems that for grain it just learned one pattern that gives it "optimal" score on average and then it just keeps using it for every single latent pixel?

File: animashit.png (86 KB, 988x547)

86 KB PNG

>>8888123
This is the setting I use with glora, again very good body, color correct. However, ~90% of the seed chance the position will have the same posing, camera angle, and even sometimes missing a limb. came/rawr doesn't do that, i've tried bat 1, bat 2 and now bat 4 already.

Either way it's kinda fun baking, i've made an extensive way to test loras and grade checkmark it.

>>8888138
The bat 4 one starts to have less of the "same position" thing i'm having problems with but regardless it still does it.

>>8888138
does kohya have a bug on amd where it does ~20s of idling after each epoch?

>>8888142
I have an amd cpu but nvidia gpu, also i'm using the ezscripts since kohya's ass gui never updated to anima.

>>8888131
oh correciton, even with a white background it does show up, it's just much more faint <https://files.catbox.moe/efykdu.png>
>>8888137
yeah I think it's caused entirely by anima(or possibly just being nvidia cosmos as a base) and it seems to be related to anything that's kind of textured but doesn't seem to exist much on flatter colors. Though, funny enough, this is the same lora, same seed, same prompt but with a different artist call(kiato is in the dataset for this lora) but doesn't seem to produce the gridding <https://files.catbox.moe/widlv5.png>
raw output <https://files.catbox.moe/djju1p.png>

>>8888123
>>8888138
oh right sorry I kind of had to do a lot of random shit.
Anyway; anima seems kind of wonky/different. It does actually get a lot more awkwardly overfit than I'd like and have been trying to do a bunch of different things to try and combat it.
Upping batch size does work to a degree but it also seems to hit reproduction a lot harder and the LR's. I've been doing a lot of random experiments/testing and so far anima does actually do a really solid job of tag differentiation for styles, which you can kind of abuse to do a sort of soft form of regularization by including a dataset from a different artist at a lower repeat count. Which, for reference, dino lora I've been posting with has dino at 3 repeats and kiato at 1 repeat(you can use whatever datasets you want, though you might want to try and make them complimentary(or contradictory) in some way).
beyond that I've also just been working on trying to tune down a batch 16(or rather, batch4grad4) config, which has been going... kind of mediocre.

File: file.png (401 KB, 3411x1909)

401 KB PNG

Gonna post this here as well but what exactly do I select to train anima? Since it's not based on SDXL or anything, what's the meta with a 4080 dataset-wise and settings wise?

I have run out of inspiration
I need to coom, but I need it to be something good

>>8888146
Out of all my tests on my end UNET only achieves better prompt positioning, I do bullshit test grading on things such as top-down, birds eye view, worms eye view, angles, dark, three times per type of prompt and checkmark "if it does x, has artifacts/clean, body issue, etc." AND also without the @tag activation.

I don't know about dims honestly I won't likely understand optimization or args you lora bakers do but i'll follow it either way, so since its 32 i'll keep it as is, filesize doesn't matter to me, quality and malleability as much as possible does.

>>8888154
All you have to do is enable the anima section

>>8888156
pov netorase handholding, the thinking (white) man's fetish

>>8888156
try this concept >>8887567

>>8888156
https://exhentai.org/s/3aabd8b072/3957151-8

you disgusting bobblehead pedo ill choke the funkopop out of you

>p*interlyfags are shitting up my vidya too now
https://files.catbox.moe/hzbt2y.jpg

https://x.com/Chief30k/status/2049937929418473803
western anime bros, we are so back

Does timestep sampling matter at all? some has sigma others sigmoid, even that russel guy's confusing lora example is in a uniform type of timestep sampling.

>>8888207
Garbage, looks way too 2d.

>>8888211
>timestep sampling
this is needed if you want to train on some noise levels more frequently than others or just change how you sample noise levels during training
picrel shows shifted uniform and shifted logit normal distributions >>8887847
logit normal distribution focuses more on the most semantically difficult timesteps (desu that's questionable) and basically never giving the model the opportunity to learn at t=1.0 and t near 0.0
uniform overfocuses at t=1.0 and rarely learns at t near 0.0
here, t=1.0 is full noise, t=0.0 is zero noise (so it's just your image)

if you're training a character or whatever with logit normal, it may look like the model converges faster but it'll probably have funny composition/anatomy and funny little details like textures

you sick bobbling child molester your head enbiggening days are numbered

what is blud even going on about

File: sigmoid.png (20 KB, 724x238)

20 KB PNG

>>8888238
You mean sigmoid (logit normal) is considered bad? ezscripts says that's the one recommeneded. so far looking at loras its sigmoid dominant with a mix of sigma, shift and only one I found in a uniform is that greg by russel to which I feel he hides something in his preview 3 lora.

>>8888242
it isn't necessarily bad, it's just different and it has the most fucking stupid flaws that nobody actually bothered to work around (or at least i think nobody did)
use whatever you like more

File: 5s.png (8 KB, 231x250)

8 KB PNG

>>8888246
I tried my best searching for more about these but considering you know more, which one is better then? I agree with you about (sigmoid) logit normal, it gives a bit of randomness on positioning/anatomy, Sigma does things better at anatomy but it has problems of the same type of positioning so i'll assume uniform is much worse.