>>109613796
>>109613796
>>109613796

use the fixed OP thread please

Is Krea or Anima better for hentai specifically?

>>109613853
*cries in illustrious*

>>109613853
go to the other thread and suggest how the models should be sorted out

>>109613868
Answer my question...

>>109613885
I did. Just not in this shitty troll bake. If you guys want more people to participate we have to do something about the trool keeping the OP stagnant and not as helpful as it could be. Also the anons here seem like bots

>>109613853
>>109613885
Krea just isn't there yet for hentai, maybe with loras in the future

>>109613890
Holy autism. Thanks for splitbaking our already slow general I guess.

>>109613908
Well you won't just fuck off and wait for the thread to get to a reasonable page or let us update the OP with relevant info. Fuck off

>>109613960
This art style is gold haha

So NovelAI v5 is broken garbage
Anima is stagnant
Krea2 is censored bad faith model

It's over for anime?

>>109614012
>So NovelAI v5 is broken garbage
Why is it broken?
>It's over for anime?
I'm happy with what we have desu

explain why you aren't posting in >>109613796

File: artist_kokaki mumose, art(...).png (1.34 MB, 832x1216)

1.34 MB PNG

>>109613971
it's a good art style.

>>109614030
>>109614029

>>109614030
Nice, Yoshizawa supremacy tho

>>109614037
>>109614029

>>109613956
>>109613960
>>109614016
>>109614030

>>109613821

>>109614029
>>109614035
>>109614040
what is this weird desperate spam?

>>109614056
Local spammer/samefag/splitbaker. Better to ignore when he's like this.

>>109614056
Quit using the schizo troll threads with outdated OP

thanks for bakering
>>109614012
im hoping the h3 image model is good since it will take references

>>109614061
You've been posting old images to bump the thread to get people to think it's the real thread. I just want to pfix the OP so it helps noobs more and to get rid of the trool

>>109614064
>im hoping the h3 image model is good since it will take references
That sounds great. I'll be trying that out as soon as it comes out, although I think having to make loras will be unavoidable. Krea has a reference lora too but it has pretty limited applications.

>>109614076
go suggest it for the OP in the other thread because that baker actually cares

>>109614081
This

>>109614063
+like
+ii ne

>>109614064
go to the fixed op bake >>109613821
I hate this troll baker that keeps stagnating info

>>109614101
i don't find you spamming and harassing every anon to be a very compelling argument for obeying you tb h

>>109614131
to be fair, the troll false flags to annoy people into staying in his threads

>>109614166
you can, it's called the pleasure of being cummed inside

>>109614166
You suck because you let the troll win. Give some life to the other thread before it's too late lest you be cursed with a drama faggot eternally

>>109614184
this isn't even the other baker (Me). fuck off

Who do I even believe anymore...

>>109614222
was her anime any good

File: nai v5 artist mixing cope test.jpg (2.49 MB, 2560x2016)

2.49 MB JPG

>>109614012
>It's over for anime?
yes

File: artist_kokaki mumose, art(...).png (1.3 MB, 1024x1024)

1.3 MB PNG

>>109614225
it was a fukkin' ride.
whether that translates to "good" or not is highly dependent on your own personal tastes. I enjoyed it, though.

We are discussing OP format in the other thread. if there is any nai users that have tips or what is the best way to sub, go suggest it there

File: ComfyUI_03195_.png (1.07 MB, 1024x1024)

1.07 MB PNG

real hollowwald translation when. the AI one i tried didnt even translate shop menus

File: file.png (1.43 MB, 1024x1440)

1.43 MB PNG

>>109613786
>>109613821

>>109614339
lame you posted here. delete and post in the other thread (or just repost). we need to get anons on board to making a better /adt/!

Good gens ITT

>>109614613
It's a very popular artist from X.

>>109613821
?

>>109613874
LOOOOL derp!

>>109613786
aren't you the same schizo that splitbakes /ldg/?

>>109614016
oeh that ones the special one

>>109614237
so glad, she finally went after her dream, heck the consequences

File: Aria wombforce9000.jpg (332 KB, 832x1216)

332 KB JPG

>>109614012
new nai learning curve is old nai learning curve
or they fucked it and have to x.5 emergency patch artists (again)
https://files.catbox.moe/h61xfm.jpg
>>109614232
this

File: Anima_06909_.png (2.56 MB, 1280x2150)

2.56 MB PNG

got inspiration from a twitter e-thot I follow

>>109615223
go back

>>109615223
a

>one out of every seven slops on V5 is wall-eyed
pfffthaha

>>109615500
it can gen transparent backgrounds natively? that's pretty neat

File: artist_kokaki mumose, art(...).png (1.07 MB, 832x1216)

1.07 MB PNG

>>109615764
Yeah, it can be a bit really wonky though and have dark splotching.
Also worth noting that, technically speaking, there's a long lost forgotten way of getting transparent backgrounds from SDXL models, too. Though I have no idea at all if it's compatible with any modern UI's anymore.

File: v5 retry.jpg (1.57 MB, 2496x1280)

1.57 MB JPG

>>109615178
Tbf the 2 tags I use are very not similar to each other and was more a thing I did a long time ago as a joke. Other artists don't result in disastrous filtered-photo/blender-assets-looking backgrounds.
I push the numbers up and got something more palatable, disregarding the background and odd chair.

>>109613640
How did you make this
Genuinely curious

File: v4.5 vs v5.jpg (611 KB, 1664x1216)

611 KB JPG

>>109615958
Today I hurt myself

>there are people who still using dinosaur model like SD1.5

File: chrome_2026-08-21-20-08-11.png (1.92 MB, 1778x919)

1.92 MB PNG

>>109616037
anon it's nai. it has metadata.
the short of it is leading with a mixed media prompt and then abusing character positioning.

>>109616065
can't believe you would do this

It's my birthday nigga

birfday girl

File: artist_kokaki mumose, art(...).png (2.65 MB, 1088x1920)

2.65 MB PNG

>>109616095
do what, exactly

>>109616793
you know what you did

>>109616065
Can you do the same thing with controlnet in NoobAI?

File: 83.jpg (318 KB, 1066x1600)

318 KB JPG

>>109616146
Happy birthdayyyy!!

File: artist_kokaki mumose, art(...).png (1.34 MB, 832x1216)

1.34 MB PNG

>>109617029
if you sketched it out? maybe? tag bleed would be a royal bitch and you'd probably have to spend a lot of time in post compositing. You're also not going to get a faux realistic car interior out of noob. Maybe with anima and whatever regional control is available on cumsoft(or at least I think? there's some form of it there) or forge couple or whatever regional prompting bullshit is available but it's still going to be a pain.
>>109616975
yeah I don't have the faintest.

File: 1775091550089041.webm (794 KB, 720x1296)

794 KB WEBM

Too many JPEGs and PNGs
Too little Webms and MP4s

..... who am i kidding, i also ran out of ideas

>>109616065
Mixed media is for troons

>>109617293
videogens works great on sloppers, they eat this stuff up instantly.

>>109617293
>wob wob wob
that's what I hear

>**Emphasis Weighting Bug Fix**
We identified and resolved a bug related to emphasis values between 0 and 1, which made some elements overwhelmingly powerful. After this fix, tag mixing should be much better. In practice, this means that your prompts will need some tweaking if it used emphases between 0 and 1, and generated images will be different as well. Furthermore, this fix is not intended to replicate the behavior of v4.5; v5 is still a different model with a completely different prompt understanding and behavior. However, it should be more intuitive now.
How tag weighting is some type of bug fix?

>>109618063
We fuckie wuckied and made 0.5% actually 500% cuz decimals hard

>>109617061
No I mean, if there's a character positioning tool like that for NoobAI. It looks really convenient.
>realistic car interior
So you could do it with NovelAI?

Is NovelAI really that much better? I'm just learning so don't troll me. I doubt I can replicate that picture locally :[
When would it a better usecase to gen locally with ComfyUI then?

https://www.reddit.com/r/StableDiffusion/comments/1vmdxzk/psa_im_the_creator_of_heretic_and_i_advise_you_to/
So it’s all just snake oil?

>>109616146
Congrats!

File: ComfyUI1_00004_.png (2.49 MB, 1440x2048)

2.49 MB PNG

>>109613786
posting this in public, had like 0.2 seconds to inspect it plz no ban

File: ComfyUI1_00006_.png (2.25 MB, 1440x2048)

2.25 MB PNG

>>109616146
happy bday, what prompts did your parents use share pls

>>109613874
>>109613956
This isn't anime styled. At least gen Konata in the anime style if you're going to post her here. Or if you want to do this alternative style, then post it in /sdg/.

Thanks for the suggestion.

File: res_mult_1.5x, 0.7__artis(...).png (1.76 MB, 1216x832)

1.76 MB PNG

>>109618722
haven't used anything regional prompting related since sd1.5 days so no clue(and back then it was very akin to html div sizing). I don't think they're anywhere near as easy/convenient, though.
>Is NovelAI really that much better?
Yes and no? I'm not exactly shilling for NAI here. v5 literally just (publicly) released two days ago and it's probably their largest jump in terms of feature set/usability ever(I would argue it's a larger jump than when they went v2>v3), but the model still has a lot of caveats.
Though yes, if your goal is to spam a bunch of characters in the same composition, NAI is probably going to be significantly better than basically anything you can do on local. Just don't expect to be able to do any weird sprawling thing with 20 characters interacting with eachother in some way.

Can't replicate my style in NAI v5 :( maybe 5.5 will fix it? :( hoping their inpainting model turns out to be better

>>109619931
>the model still has a lot of caveats.
NTA, but curious and out of the loop. What are those caveats?
>Just don't expect to be able to do any weird sprawling thing with 20 characters interacting with eachother in some way.
Possible victory for local, looking at the gen in comfyui_guide_1girl rentry, maybe NAI has the tools to do the same though

>>109619959
>>109618722
NAI, like any model, is a tool. I was and still am stupidly happy with 4.5 full, I found a style that actually worked. Then I burned through half my Opus credits trying to get it working in v5 and just... couldn't. Can't even replicate it locally, I got lucky and 4.5 got some secret sauce tuning for it. Tried training LoRAs of it for Anima too. Nothing. It doesn't get it. So yeah, AI progress isn't some clean upward line.

Find your style and get lucky once, and you might be cursed to never find it again anywhere else.

>>109620010
>Tried training LoRAs of it for Anima too. Nothing. It doesn't get it.
Might've had better success for IL

File: res_mult_1.5x, 0.7__artis(...).png (1.34 MB, 832x1216)

1.34 MB PNG

>>109619965
>What are those caveats?
Aesthetically it is not the best model. And prompting certain types of artists can degrade compositions(and alternatively you may need/want to prompt other types of artists to attempt to improve certain compositions).
Also what they're pulling with the opus limits is pretty scumbag/anti-consumer behavior.
>Possible victory for local, looking at the gen in comfyui_guide_1girl rentry, maybe NAI has the tools to do the same though
Pretty sure that sort of thing took monstrous effort(and still looks a bit wonky) on local. That said, it'd also still be pretty involved and a huge pain in the ass to do on nai. Probably would be a good bit easier to do though since they have an actual inpaint model and you can try and do things a lot more piece by piece(you can do that on local, too, just probably won't go quite as well).
Meanwhile though, the bocchi car wash thing was effectively one shot t2i and probably took less than 10 variations/adjustments to get the output.

>>109620053
Thanks for the deets. I quite like NAIV5 aesthetics so far, though I don't have much to compare it to.

>>109620091
Aesthetics is independent of whether a model is advanced or not.
Using NAIv5 for example, I'm starting to feel put off by the 32ch VAE. I think that kind of VAE is only suited to certain situations.

>>109613891
I tried, it's partly successful: https://huggingface.co/quarterturn/krea2-danbooru-1024-40k
It was trained on 40K danbooru 1024 explicit and questionable-rated imgages. it improves character recognition and h-scene comprehension, but krea2 is still there fighting against you on the poses you want.
I trained until step 15000. From 12000 to 15000 loss flatlined, so I think it's done at the last epoch. YMMV no refunds.

Finetuning these models year after year has gotta be rough, since so much of the influence comes down to tagging and the dataset, not the model itself. One year where artists worldwide start uploading a certain style to danbooru and the whole dataset gets skewed

>>109620193
Do you train it on your computer or do you rent a GPU? If you rent, how much money? How do you tag so many images automatically? I'm guessing most of it comes down to a leap of faith

>>109620270
I used my own GPU, a 4090D 48GB. The images were captioned by combining Danbooru 1024 image set with the Danbooru 2026 caption set which was fed into a custom python script calling minimax-m3 via openrouter. This way, m3 could be told what characters were in each image, and what they look like, and was able to very accurately describe where they were in the image and what they were doing. This was the resulting dataset which was used to train the lora: https://huggingface.co/datasets/quarterturn/danbooru-1024-eq-captioned

File: CrTd0SkW8AAh3G4-1027312207.jpg (56 KB, 379x680)

56 KB JPG

>>109620308
>used my own GPU, a 4090D 48GB

>>109620337
$3K doesn't seem so expensive now, compared to where 5090s and 6000 Pros are at. I got it from C2 Computer. I've run it for weeks at a time at 100% full power limit, it's solid. I would buy another but I really want more VRAM on a single GPU.

>>109620361
>spent $3k on a 5000 ada when the 5090 dropped
>5090s are now $5k+

File: konata-lasershow-1.mp4 (1.81 MB, 640x640)

1.81 MB MP4

>>109620622
32GB isn't enough for the job. I tried like five times to get an MSRP 5090 at bestbuy too. No luck.

File: 62829777-cb8e-4d8e-a862-3(...).png (1.21 MB, 2048x3072)

1.21 MB PNG

ULTRABASED:
https://civitai.red/models/1217645/sih
ΣΙΗ ILLUSTRIOUS/NOOB FINETUNE UPDATED

Please:

del "anima-aesthetic-v1.0.safetensors" "anima-base-v1.0.safetensors" "anima-turbo-v1.0.safetensors"

File: c2af7167-6c97-4a73-9309-c(...).png (924 KB, 1564x2502)

924 KB PNG

ULTRACRINGE WTF IS THIS MENTAL ILLNESS??
https://civitai.red/models/2880480/anima-38b
ANIMA 3.8B
If every checkpoint maker had donated their money to TDrusell instead of spending it inhis personally failed, half baked project, we would have Anima 2 by now.

>>109621108
It looks more broken and slop than I could have imagined

>>109621080
It looks more fixed and KINO than I could have imagined

>>109621108
Just train proper CN for Anima, retards.

File: 00021-3659253567.png (1.03 MB, 832x1216)

1.03 MB PNG

my hands felt just like two balloons

>>109621080
Interdasting

>>109621245
WAI there some 1girls smiling tho THO?

>>109621324
a happy ending

>>109621179
No, better to get 5 minutes of fame from a model nobody will ever use once or twice, before they realize it takes 3x longer to render an underbaked result... for something you could've just done with inpainting anyway

File: Twitter_post_-_NAID_V5_24(...).jpg (213 KB, 1200x675)

213 KB JPG

how many local gens?

>>109621484
probably the same amount but with videos kek

>>109621484
Al lot more, but 12.6M is also impressive desu
What does the average gen cost?

>>109619931
Most of the things I see in NovelAI website just look bad, thats why I didnt pay much attention to it

So NAI can generate that car interior and all that. Im surprised

I guess im gonna try later to replicate it with my NoobAI workflow that im building (I moved away from Anime cause it has zero tools)

>>109621484
most of those are just tinkering gens. tag weighting might as well be homeopathy at this point, less severe than Anima but still funny people pay for it, anime screenshot tag is fryed

>>109621245
Cute AIgold.

>>109621245
How long did it take to gen all of those?

>>109621245
I bet all of them have a full image they are cropped from.

>>109621503
Well with the $25 opus sub if you consistently gen 190 (recharge rate of V5 usage limit for $25 sub) 1MP 23 steps images per day for 30 days straight it can be $0.0043 per image (not counting the 1.5k buffer), though not everyone will use up all given capacity.
If you don't sub, then buy the $13.99 anlas pack, 384.6 default gens costs $0.036 each, basically laughing at non-subscribers.
if you get the $10 sub just for the pack discount, which only makes sense if you only gen >1MP images (since large images aren't "free" in the opus sub) and will only gen small amounts across several months, and grab the $10.99 pack twice, so a total of $31.98, 769.2 default images will cost $0.0285 each, going down only if you buy massive amounts of anlas but nowhere as efficient as subbing for real. Anlas is really meant for "real users" to use on Precise Reference, VT encoding, shit like that.
Also, the 1 anlas (from 1MP) 2x upscale actually doesn't look terrible? I guess if you need 4MP you can gen 1MP with the opus sub and upscale those, if you still need to inpaint, then crop 1MP out of it and inpaint that.

File: 1787427003574.png (2.93 MB, 1408x2176)

2.93 MB PNG

[@artist1:@artist2:0.5] actually works for Anima artists mixing in Forge Neo. It's more stable than (@artist1), (@artist2).

File: ComfyUI_temp_xnglz_00005_.png (782 KB, 736x1024)

782 KB PNG

>>109621783
ty do they use this mix?
>>109621791
about $3,50
>>109621820
nope :)
https://files.catbox.moe/w7ikcd.png

>>109621922
if all are subbed they net 18 mil a year, but since the non-subbed gens are almost 10x in price, that would be a brutal moneymaker
A third of the gens costing 3ct would net 44 million

>>109621957
>(@artist1), (@artist2
>>109621699
tag weighting without clip might as well be homeopathy

Nai 5.5 when?
V5full=big undercover flop

>>109621699
they actually ""fixed"" weighting a bit last night. It's not perfect but lowering artist weight below 1 actually does something now and you can tune out artists that were overpowering everything else.

>>109621990
but how do you know how many subbers there are

>>109622484
my logic, they say they gen 12.6m picas a month
if all of those pics generated are subbed gens they would all cost 0,4ct a piece, regardless if there are x or y amount of users generating those 12.6m

The figure of '30% isn't subbed, ergo pays the more expensive price of 3ct a gen = (4mil x 3ct x 12 months) is just a guess that 1/3rd isn't a sub that uses the site

NAIcord has roughly the same amount of people as the ComfyUI cord. They've got a big Japanese userbase and they're clearly making good money off this. That said, they're also doing a lot of damage control lately: people are complaining that the anime screenshot tag isn't working properly, meanwhile the image posting thread keeps getting flooded with anime screenshots anyway. At the end of the day it's a company, so hopefully they actually release 5.5. A model that only comes out once a year can't just be treated as a one shot release and then left untouched until the next year rolls around.

File: 1787432705451941.jpg (1.12 MB, 2688x1803)

1.12 MB JPG

>>109622639
Long femur, long forearm. The image content stopped way before the canvas edge, cropped it to 2:3 and it made way more sense. AI just can't handle aspect ratios outside 2:3/3:2/1:1, one can always dream, but something always ends up elongated.

File: res_mult_1.5x, very aesth(...).png (1.6 MB, 832x1216)

1.6 MB PNG

>>109621645
>Most of the things I see in NovelAI website just look bad
keep in mind NAI only shows off images without any artists prompts and without any named characters, so of course most of it looks pretty lackluster.
Not saying it doesn't still struggle with aesthetics but it's not that bad.
>So NAI can generate that car interior and all that. Im surprised
It can do an amount of photorealistic. It often garbles itself though unless you keep it fairly simple. I had posted a few other examples last thread
>>109607355
>>109607555
>>109607689

File: file.png (47 KB, 547x255)

47 KB PNG

>>109615958
I found out that the weights emphasis were fucked yesterday before the crash. It didn't occur to me to retest it until now.
Now
>0.75::one::, two
werks just as expected.

>>109618063
Their "tag weighting" seems to be implemented through LoRA or embedding injections instead of the text encoder, there's no text encoder issue to fix. What needs fixing is the "anime screenshot" style, because it overpowers whatever other style you're going for.

>>109622809
>werks just as expected
Still broken. Though v4.5 also had broken tag weights, at least here "expert shading" does something.

>>109622800
You have a very trooon style

File: Pink6.jpg (1.33 MB, 1472x2048)

1.33 MB JPG

>>109622772
But now you can't see her legs...
>>109622800
I am really enjoying the mixed media angle
>>109622808
that's a horse

>>109622893
Woaw. All that detail makes it look really nice.

File: IMG_1112-2.jpg (2.85 MB, 2592x1944)

2.85 MB JPG

>>109622893
>I am really enjoying the mixed media angle
The more fun part is taking an actual photo and inpainting over it to inset a character. Again, it's not perfect at it, but it can actually orientate to surfaces/perspective. Of course though, this actually takes a bit of effort to 1.crop the photo to inpaint(to be at a compatible resolution) and 2. doing a proper composite back to the original, especially since anything in the inpaint area is going to get fucked and you'll need/want to get rid of anything that isn't the subject.
shadows are still pretty much beyond it, though.

>>109613786
>NEW
NovelAI v5(Paid)(SOTA): https://novelai.net/updates?id=naid-v5

>>109623011
I am using another anon's upscaling
https://files.catbox.moe/nt7jo3.png
and I agree it's wonderful

>>109623105
pre-ai I l o v e d the russian anime girls in dystopian depressive setting route
with ai I didn't really try but good suggestion

>>109623105
?? You can't do that in Comfy?
Is local so behind?

What is local good for then? Just 1girl in typical illustration or what? I don't get it

>>109623246
one can but I'd argue finking can be half the journey.

File: 1787439241995.jpg (252 KB, 1408x2176)

252 KB JPG

OneObsession Anima v3

you can definitely do that on krita, and it uses comfy as a backend

>>109623246
It's not that local can't it's that it takes a monumental more amount of effort and is less flexible. The bonus here is that you're doing it with a, natively, chinese cartoon oriented image model that can also function in a context model capacity(you can inpaint a random image without calling any styling tags and it will style the inpainted area to match the rest of the image).
Inpainting(and how much easier/more effective it is) is legitimately is one of the best selling points of NAI.

>>109623288
oeh I like that
the zipper, laced side and the face are nice, the blue is original if it was your choice

>>109623288
>that zipper and o rings
Did you inpaint that? There's no way Anima can do that kind of detail on its own.

>>109623323
She's feeling if they are ripe for harvest

File: 1787440168493.jpg (1.72 MB, 2880x3840)

1.72 MB JPG

>>109623370
> Did you inpaint that?
No. It's anima+LAX's Tile&Repair CN+HiresFix and Adetailer. This is the reference.

>>109623291
Nah Krita's trash, not sure who it's made for. You have to dig into the config window to change inpaint settings instead of having them easily accessible like Forge or Invoke.
Also Krita doesn't let you swap, steps, samplers, cfg, loras or checkpoints on the fly, it saves everything under "styles" per checkpoint, as if nobody uses multiple styles in one gen. It's bizarre, not even sure Acyl uses his own program for real.

File: compressed_1787440405426.jpg (35 KB, 300x385)

35 KB JPG

>>109623401

>>109623401
Share the workflow! I could never get controlnet tiles working in Forge, let alone the LAX one in ComfyUI. A screenshot would do if you're too prude to hand over the actual file.

>>109623452
Sorry, I use Forge Neo.

>>109623401
neat, the gen looks better than the pica

File: file.png (30 KB, 542x140)

30 KB PNG

>>109621922 (me)
>2x upscale
oh shit that's also a new thing, I vaguely remember using the old one months ago once and it was this shitty and useless 4x thing that made everything blobby

>>109623511
>She's going to eat me
Finally!

File: res_mult_1.5x, very aesth(...).png (888 KB, 832x1216)

888 KB PNG

>>109623513
it's just reskinned waifu2x. nothing really worth caring about. The "real" improved upscale is hidden behind "MAX" enhance but that costs anlas and is a bit too finnicky for regular use.

File: 1773830715758417.png (1.93 MB, 1024x1024)

1.93 MB PNG

>>109623373
they seem ripe for harvesting, now she has to taste them
https://files.catbox.moe/fw3raf.png

>>109623535
the new enhance is much better alongside its max function, but i wish it didn't brighten the images, it also does it in a weird way. feels like it 'flattens' the image

>>109623535
waifu2x run slow on my outdated potato but yeah it should've been a thing over a year ago

File: Comfrea2_00008_.png (2.29 MB, 944x1672)

2.29 MB PNG

>>109621184
I can't explain, you would not understand. This is not how I am.

>>109621484
>The best tools of the hobby are behind a shitty corpo that charges 25 bucks a month
>People unironically paying for it anyway
Grim and sad as fuck. I thought it was just some service for casual people that didnt want to dive into comfy or complex workflows, not that it was actually good.

So, all those anime coding autists, specially Chinese ones, beaten by a shitty corpo. Sad.

>>109623690
> The best tools of the hobby are behind a shitty corpo
No CN, no Adetailer. The best tools my ass.

>>109623714
Thing is, you don't need XinsirNet if the model already has good prompt adherence and good img2img understanding and interpretation.
Also, you can actually automate NAI's adetailer through the NAI webUI extension or Bedovy's ComfyUI NAID pack.

>>109623714
>shitty corpo
Thank the hundreds of disorganized local finetuners who dump datasets, resources, time and money into failbakes like that one >>109621108
You know what's actually most coveted, right? Good quality, well tagged datasets. So what happens if I tell you that among local finetuners, only a tiny minority ever share those, because that's the part that actually has value, and all they really do is hand you their shitty closed model for free?

>>109623790
3.8b was trained with 40 hours on a SINGLE 4090
There are no effective resources that went into that. It's just some retard doing dumb stupid shit on their own.

0.7::artist:dino (dinoartforame)::, 0.9::artist:ciloranko::, 0.7::artist:deal360acv::, 0.7::artist:wepwawet::, 1::artist:ootorii bisetsu::, 0.5::chromatic aberration::, gritty, faux traditional media, painting (medium),1.5::masterpiece::, 1.5::highly detailed::, 1.4::anime coloring, anime style, anime art::, 1.4::cinematic composition::, 1.3::dramatic chiaroscuro::, 1.3::luminous modern colors::, detailed face, detailed clothing, realistic anatomy, intricate textures,

What causes this?

>>109623803
What causes this?
Peak larping behavior. Imagine using a music gen tool and larping with tags and weights like 1.4::beethoven::, 0.8::mozart::, 3::nirvana::, very intricate guitars, detailed drums, like stacking those numbers actually makes you a composer.
NAI users are some of the biggest larpers to ever exist, and they finally found a tool that lets them larp freely and openly.

>>109623803
Where's tiandingdong?

>>109623819
Weird way to spell autism

>>109623801
Local is just retardeds doing their own thing. If NAI's team split off like that too, everyone making their own solo product instead of one cohesive release, NAI wouldn't last.

File: res_mult_1.5x, very aesth(...).png (1.31 MB, 1216x832)

1.31 MB PNG

someone seems rather upset.
>>109623803
also why are you making up half those tags. the metadata is right there, assuming you're not a dumbass and have the catbox extension you can just rightclick the download button and display it.
Unless you're talking about something somewhere from someone that isn't me who copied half the style prompt and yolo'd the rest of it.

>>109623859
>If a company wasn't a company anymore
Durrr durr

File: P2 Miko.jpg (273 KB, 832x1216)

273 KB JPG

>>109623882
Sniff sniff
Jussy detected

I'd make some Krea2 loras if I knew what a decent dataset looked like

>>109618567
What model is this?

If I decide to use Forge Neo over ComfyUI (to get started,) do I have to take into consideration using it in a container like ComfyUI? I know the guide recommends using docker for ComfyUI.
I plan on using WAI Illustrious SDXL NSFW. I have an RX 6600 (8GB). Is there anything that would be useful for me to know before preceding to ensure that everything runs smoothly?

File: 1769801917390778.png (1.77 MB, 1024x1024)

1.77 MB PNG

>>109623899
sniff
she deserves better

>>109624237
nai v5, metadata in the image is fucked up though. i can catbox the original if needed

>>109624349
>nai v5, metadata in the image is fucked up though. i can catbox the original if needed
Please do, love that aliased painterly pseudo-lowpoly style

>>109624342
All python webslop should be contained

>>109624384
Sounds good then.
Pardon if this is a retarded question, but what would be the suitable way to contain Neo Forge?
As for ComfyUI, I see this is mentioned in the generation guide, but I was wondering if there were any specific things to take into consideration with it?
https://github.com/YanWenKun/ComfyUI-Docker

File: 1755958265281408.png (1.72 MB, 1024x1024)

1.72 MB PNG

>>109624360
https://files.catbox.moe/5mmuht.png

>>109624458
Thank you hero

>>109624412
Oh, I am retarded. I forgot to mention that I was on Arch Linux.

I have a question for the people here. What programs do AI artists with at least 300+ paid subscribers on Patreon or other sponsorship platforms use?

I’ve been serializing a dialogue-driven, medium-length story series as a hobby on a small Discord server. The reception was pretty good, so I recently opened a Patreon. I only used an early local model back in 2023 and have been using NovelAI (NAI) ever since. However, moving beyond a hobby to regularly delivering monthly content to subscribers made me realize how expensive it actually is.

Because I’m a perfectionist, I always try to push for total perfection—for a 30 to 40-page comic with dialogue, I end up generating around 500 to 1,000 sample images. Generating 500 images burns through all my Opus subscription credits. I’ve spent $140 on extra credits this month alone, which feels like a total waste.

Are there any exclusive communities, Discord servers, or Telegram chat rooms specifically for AI NSFW creators? Does everyone use local models? My PC specs aren't great. I’m really curious about what top-tier creators with thousands of subscribers—like Potion or helltoyou—are using anons, stop messing around and answer me please. The previous thread was closed so I couldn't post my reply. If you leave a quality response, Donald J. Trump and J.D. Vance will bless you all, my fellow Americans.

>>109624412
Docker images are the easiest unless you know podman

>>109623790
>Good quality, well tagged datasets.
anima was trained on the gelbooru, ye-pop, and deviantart datasets on huggingface. literally just download them and you have anima's dataset. the datasets aren't the problem.

>>109624210
Beautiful, sexy Konata AIgold. Can you gen some Kagami too?

If you got 16GB+ vram, what models do you use now?
>Illustrious
>Anima
>Krea 2
>Flux 2 Klein
>Other

>>109625373
16gb included or only over 16gb?

File: 00008-434502327.png (926 KB, 1216x832)

926 KB PNG

>>109625373
didn't you post this on /napt/ already sis?

>>109625099
Hot shape.

>>109625409
16gb included

>>109625419
I want multiple answers, a survey for my interest basically

>>109625373
Illustrious-family checkpoints
Krea 2
H3

>>109624210
This is so good.

>>109625099
>>109625446
this
delicious tum fat

>>109625849
Rare Anemone in the wild.
>>109625980
Blue board, anon.

>>109625980
Well fuck, I did that by accident twice while cross posting. Hopefully the jannies and advertisers don't find out.

>>109625686
nice miku

File: Krea2_turbo_00233_.png (1.21 MB, 1448x1448)

1.21 MB PNG

If anime were displayed on a 1970s vector computer display it would look like this

File: 2026-08-23-174240_anima_b(...).png (3.02 MB, 1024x1536)

3.02 MB PNG

>>109626052
Anemone has one of my favorite designs in anime

>>109625373
Localchads use and try everything

In the process of setting up ComfyUI docker image. Running out of space in root while the image is being pulled, and it fails. Will setting the docker data location fuck with the security any? I was trying to set it up in a specific folder but got this error thrown:
>docker: Error response from daemon: failed to mount /tmp/containerd-mount3767912742: mount source: "overlay", target: "/tmp/containerd-mount3767912742", fstype: overlay, flags: 0, data: "workdir=/home/anon/containerd-data/io.containerd.snapshotter.v1.overlayfs/snapshots/2/work,upperdir=/home/anon/containerd-data/io.containerd.snapshotter.v1.overlayfs/snapshots/2/fs,lowerdir=/home/anon/containerd-data/io.containerd.snapshotter.v1.overlayfs/snapshots/1/fs", err: no such device
I think I would have had to re-mount my /home with a certain option to work, but I needed to do a bit more research on it so I don't fuck anything up. I read about remounting it with those options potentially messing up the file permissions.

>>109628580
dinner

>>109625373
You get honest with yourself and take a hard look at your digital surroundings.

-LAX keeps churning out broken ControlNet models.

-CivitAI is drowning in bad faith buzz earning bait slop.

-The handful of finetuners active in noobcord are making solo finetunes out of spite toward Anima.

-Anima the SoTA model, 1 year outdated daraset, architecture broken, nobody cam build on top of it.

-tdrussell walked away from the scene entirely(I would do the same if I were in his shoes)

-renting compute for training ends up costing more in the long run than it's worth, since GPU, RAM, and storage prices keep climbing.

-new local models keep arriving nearly untrainable(krea 2, klein, anima), wrapped in a bad faith safety filter.

So....

That's the moment you open NovelAI.net and see they dropped a new model 2 days ago, 13 million gens already, It combines the prompt adherence of GPT Image 2 with the quality of Flux VAE2, and it actually has current, up to date knowledge of styles and characters, something local models are a year and two months behind on
.
NAI is our company, anon. We belong to NAI. Any childish attempt to stay "independent" by clinging to local models is just a waste of time at this point.

>>109629535
Ironic how Civitai's reward system was supposed to raise quality and it made everything race to the bottom instead.

File: 00023-2767618789.png (1.04 MB, 832x1216)

1.04 MB PNG

Welcome home, good Hunter. What is it you desire?

>>109625099
what is res_mult:1.5x?

>>109629535
why did tdrussell leave?

>>109629706
Go ask him on his Hugging Face page. His silence alone might be the answer you're looking for.

>>109625207
Omg how cute!

>>109624614
At your volume, local gen isn't optional. At 500 to 1,000 images per issue, paying per image will always cost too much. Creators who keep up a monthly Patreon schedule almost always run Stable Diffusion, Illustrious, or Anima models on their own PC through ComfyUI or Forge. Once the model is on your machine, making images is free. You only pay for power and time but I think your PC is the real problem to fix, if you can't run it locally yet, rent a GPU instead, paying for rental time, it's way cheaper than per image pricing at your volume.

File: res_mult_1.5x, very aesth(...).png (1.9 MB, 1216x832)

1.9 MB PNG

>>109629829
something with no real documentation because the devs feel it's more likely to ruin output than anything positive(read: it has a minimum skill floor)
the "default" is something like 2.1x or something around there. It sort of? Tells the model to scale the output resolution(relative to the dimensions). Think of it like an attempt at a pseudo supersampling type thing in image gen. It's weird and kind of hard to explain.
the tl;dr is that lowering it should give you slightly crunchier/more textured outputs and increasing it with smooth things out a bit.
I wouldn't really worry about it much unless you really want to experiment with it.

File: cui_peach_0036.jpg (918 KB, 1600x2048)

918 KB JPG

>>109625373
I have only 8 gigs and run everything there except Flux

>>109630036
And why do you habe that type of knoweldge that isn't in the novelai docs?

>>109630066
The bane of vramlets

File: Discord_2026-08-23-16-56-47.png (44 KB, 408x475)

44 KB PNG

>>109630201
there are other reasons but just because it's not in documentation doesn't mean they haven't publicly talked about it on their 'cord.

>>109630036
>>109630213
thanks, i see, it definitely makes things look different between 1-4

>>109630213
so a different syntax for lowres/hires, I guess I'll try it. I wish they'd make NAI 5.5.
Can you shill /adt/ there? We need more NovelAI posters.

>>109630435
I hope they make 5.5, too. In my opinion, 5 is a total dud that only creates deepried slop.
I could make 4 work so it surprised me when they released 4.5 due to community backlash. And 4.5 was a massive upgrade.

But 5 is a sidegrade. It creates complex images, it understands complex prompts in natural language, but no matter what you do, the style looks shit.

>>109629597
They just need to add a system where if your model gets downvoted enough you get banned and have to pay $200 out of pocket.

File: 00024-1809749201.png (948 KB, 768x1264)

948 KB PNG

thick wolf thighs

File: nai v3 mama samples.jpg (1.68 MB, 2496x1216)

1.68 MB JPG

Anyone have any idea which artist are similar to this? Prompting "mama" in NAI V3 gives this kind of appearance. I tried searching for artist names beginning with mama but they ain't it.

>>109628580
Hows your Controlnet so good

>>109629535
>krea 2
>untrainable
Did I miss something or ...?

>>109630736
Cute
>>109630550
It feels overfitted to complex prompts, and from what I've read on the NAI cord, it also lost track of characters it already knew in v4.5. Problem is the NAI cord is full of sloppers getting fooled by complex prmpts or glitched mixed media garbage like this guy >109630036, so good luck getting them to actually train a v5.5. Wild for a hobby that moves this fast.

>>109629535
I haven't had much trouble making loras for Anima and I've heard baking for Krea is relatively easy too

>>109630036
>>109625099
Give me back my v4.5 style faggot! Your 32ch VAE SUCKS if I can' portray my style from model to another model

>>109625400
This also is cute

>>109628882
Does anyone have any ideas about this one?

>>109631417
>>109628882
Your error is a filesystem problem. Docker uses "overlay2" for storage. This needs a normal filesystem like ext4 or xfs. Some filesystems do not work well with overlay2. Examples are tmpfs, NFS, ZFS, and eCryptfs. Also, if your /home has the "nodev" mount option, overlay2 will fail with "no such device." This is probably your problem, not permissions. Maybe check your mount options first and look for "nodev" in the list. You do not need to remount /home you ca move Docker's data folder to a different location. And moving Docker's folder does not change file permissions inside it. Docker still controls that. Pick a normal local disk folder, not network storage (NFS/SMB), and not a folder with "noexec." Then you are safe.

File: 8cd0fe95-658f-40dd-b9d4-5(...).png (1.83 MB, 1024x1024)

1.83 MB PNG

Most guns are still kinda fucked up.

>>109631945
>>109631690
Consider this: 2D woman completes us, and not merely because of the "animu waifu :3333" meme. There is something ontological at work. The luminous 2D anime chibi girl requires the 3D reality of men to actualize her own existence, to gain perspective and become whole through relation to us. In turn, we require a realm defined by tenderness, cuteness, care, and safety in order to have the strength to confront the harsh reality of the 3D world, it is a mutual perfect dependency.

File: 1757211974126687.png (1.24 MB, 1024x1024)

1.24 MB PNG

>>109632002
two halves severed along the axis of dimensionality
yet forever in the state of attraction
anon has done the metaphysics
dangerously based

File: Big fans of big green dudes.jpg (603 KB, 1088x1920)

603 KB JPG

>>109631265
Tanks!

>>109629535
>new local models keep arriving nearly untrainable(krea 2, klein, anima)
>untrainable
This is the dumbest meme I've ever seen in this whole community and I mean that. How were these models trained in the first place, if they are so broken as to be untrainable? You create a model by TRAINING IT dumbfuck. If the model is good then by definition it is trainable because IT WAS TRAINED IN ORDER TO MAKE IT. You are looking at a thing created by a training process, and saying "but actually it's untrainable" like that makes any sense at all.

>>109629535
>krea 2
>untrainable
It's the best and fastest model for training anything right now you stupid shill. The real problem is that it’s a heavy model so most people can’t run it. Fewer users means fewer people make fine tunes for it. And because it’s heavy trainers have to lower batch size which hurts the quality of the output. Trump and MAGAtards killed this hobby

>>109619931
>I would argue it's a larger jump than when they went v2>v3
I like v5 a lot but honestly I don't agree, the jump from 1.5 models to SDXL and v3 specifically was the biggest anime model leap ever. 1.5 was barely usable for anything other than basic 1girl slop unless you use heavily overfit loras and controlnet. Everything after feels like incremental improvements on already usable stuff
Honestly it's kinda weird seeing you praise v5 so much because that means you were genuinely unimpressed by v3 back then and not just blindly hating on saas.

>>109632928
i'm platinum hard

>>109631847
>>109631919
They look pretty good here desu.

File: res_mult_1.5x, very aesth(...).png (2.04 MB, 1216x832)

2.04 MB PNG

>>109632951
v5 has a lot of weird random tech that actually makes it interesting. If you thought being able to do random NL on anima was cool and neat, v5 does it about 10 times better. With every other model when prompting characters I'd expect that, if they're clothed, their random bits and bobs and whatever will usually be a bit off and need correction. v5 actually manages to get things right the vast majority of the time even with gachashit tier characters with dumb overly intricate designs. Transparent backgrounds are also extremely huge. There's also the even bigger elephant in the room with it being not just being able to handle runes, but can also just plainly be prompted in whatever language you want and it will just work. Also also, the inpainting model is pretty nuts.
>because that means you were genuinely unimpressed by v3 back then and not just blindly hating on saas.
I really wasn't. SAAS is still cancer, obviously, but v3 just really was not the crazy huge leap people like to pretend it was. People like to think back on SDXL and go "oh yeah that was so huge!" except NAI was locked to 1mp on a 4ch VAE with no recourse and no controlnet options. If it couldn't do the character you wanted you were SOL, if you wanted to have characters do the sexy sex you were still in shoggoth adjacent territory and let's not forget that without SMEA it looked like trash and SMEA isn't even an actual correct implementation - just an "oops this doesn't work correctly but it works incorrectly in a good way?" implementation.
The model also aged like milk(pdv6 came out 3 months later and did sex things significantly better lmao). If NAI dropped it from their service and put it up on huggingface the only actual useful thing out of it would be the inpainting model and even that probably wouldn't be as useful as people want to believe it would be.

>>109633123
i'm a local guy but i'm impressed by how consistently ahead of the curve NAI has remained

>>109633139
(which is interesting because their text models have always seemed to be lagging behind)

The way things are going NAI’s gonna go bankrupt in the next decade from GPU rental costs alone lol. Being ahead of the curve just means you’re one foot in the grave sooner than everyone else

File: res_mult_1.5x, very aesth(...).png (1.56 MB, 832x1216)

1.56 MB PNG

>>109633139
they were only sort of ahead of the curve with v3 for a few months, technically not even a year if you want to argue the "actual anime model" standpoint(illustrious was trained in april? I think 2024, then released in september 2024, while v3 launched october 2023). v4 and v4.5 were legitimately horrible models that took two steps forward and five steps back,
even v5 is only really "good" for being an uncensored chinese cartoon model that includes artists. GPT/google completely destroy it in terms of being a context model. It also still has issues in that it's pretty stylistically inconsistent and is too reliant on prompting artist X to do subject matter Y.
Like, I'm playing around with it now because it does a lot of cool shit really easily and I'm a hobbyist. But I wouldn't suggest or recommend anyone actually pay for it unless they're ESL as fuck and have legitimate problems prompting in English, really want to play around with the inpaint model or REALLY need recent gachashit character consistency and don't want to deal with character lora slop.

File: 379476a1-b5fc-42ee-8490-b(...).png (1.82 MB, 1216x832)

1.82 MB PNG

>>109633100
I cherry picked those out of a pile of trash. The AK is a bit better on average.

>>109632629
Why no Anima updates like they were with illu? Why anima base still the only stable model? Huh? Faggot.

>>109633232
I think they have their own GPUs, but AI companies are going to get crushed by GPU costs and competition. For NAI, it might become impossible to keep improving the quality of their service without raising the monthly fee. The technology just keeps getting more and more bloated, so every improvement comes with increasingly insane costs. At the end of the day, AI is just a very well decorated scam scheme.

>>109633393
>AI is just a very well decorated scam scheme.
I was skeptical until Fable dropped. There's no doubt in my mind anymore that Fable-tier models multiply coding productivity and codebase robustness.