Confused edition

Previous Thread: >>>>8660637

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

>RESOURCES
Wiki: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki | https://comfyanonymous.github.io/ComfyUI_examples
Training: https://rentry.org/59xed3 | https://github.com/derrian-distro/LoRA_Easy_Training_Scripts | https://github.com/bmaltais/kohya_ss | https://github.com/Nerogar/OneTrainer
Tags: https://danbooru.donmai.us/wiki_pages/tag_groups | https://danbooru.donmai.us/related_tag
ControlNet: https://rentry.org/dummycontrolnet | https://civitai.com/models/136070
IOPaint (LamaCleaner): https://www.iopaint.com/install
Upscalers: https://openmodeldb.info
Booru: https://aibooru.online
4chanX Catbox/NAI prompt userscript: https://rentry.org/hdgcb
Illustrious-related: https://rentry.org/illustrious_loras_n_stuff
Useful Nodes/Extensions: https://rentry.org/8csaevw5

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

Again, no idea what happened to the last thread but who cares at this point

>>8663850
Nice job with this gen, the expression turned out really well. How do you prompt that kind of mouth?

ero manga is a thing

I wish there was a way to easily prompt wide shots that contain the ENTIRE character. It seems like all the models I'm using are so biased towards portraits and cowboy shots that the top of a characters head or their feet are slways cutoff. Especially in landscape aspect ratios ;_;

>>8663873
>>8663873
Have you tried to prompt what you want to see? but yeah, if you are not getting what you want, you should change model

https://files.catbox.moe/miosrh.png

>>8663873
That's still happening even with "wide shot" in the prompt? Surprising but if so have you tried adding "full body" or "full-length portrait" to the prompt as well?

>>8663878
>>8663886
I don't have any good examples to show, but I think the problem is that I'm prompting every article of clothing and a lot of specific body details of the subject. Like when I'm prompting 1girl, naked (only 2 tags) i can get the kind of compositions i want, but i want to use half a dozen tags describing the character and a dozen describing their outfit. It seems to just naturally pull in focus tighter than i want.

I know the solution is controlnet and/or inpainting. But i don't know how to do those on comfy.

DAILY MODEL WATCH

Here's a test of new models compared to WAI.

One of the new models today was https://civitai.com/models/1784945/bondiment , which is a mix of MiaoMiao Harem and One obsession. Looking at the models I decided to test those two since I haven't seen anyone talk about them and apparently they're very popular on CivitAI?

https://civitai.com/models/934764/miaomiao-harem
https://civitai.com/models/1318945/one-obsession

Both of these seem to be anime-influenced 2.5D models. They seem to be fine quality models with some example images on CivitAI showing they can have a more anime look with the right prompts.

Another new one is https://civitai.com/models/1784563/rurudo-style-finetune , which is a chinese model intended to emulate the style of an artist named "rurudo." Surprising to see this being done with an entire model instead of a LORA. Still it's a soft color anime look if that's what you're looking for.

Lastly, as a possible horrible look into the future, https://civitai.com/models/1785579/pony-peoples-works-base-noobai "Pony: People's Works" is a chinese model deliberately trained with AI-generated images published by users on CivitAI, mostly PonyV6 images. It seems to want to fill the image with a lot of clutter, and may have more of the anatomy issues that PonyV6 had. An interesting experiment perhaps but not something I'd recommend using seriously,

Full res test: https://files.catbox.moe/ienikp.png

>>8663908
>Surprising to see this being done with an entire model instead of a LORA.
Pretty sure we have like 5 or 6 anons doing that here. The question is - why upload the full checkpoint instead of a locon extract.

>>8663850
>no idea what happened to the last thread
got dumped off the board again >>8663797

>>8663873
>>8663891
I think what you need to prompt is "barefoot" or the kind of shoes the want (even something as basic as "brown shoes").
Here's a simple guide/workflow for ControlNet: https://rentry.org/comfyui_guide_1girl#controlnet-pose-transfer

>>8663865
I like this. Good job.
>>8663864
I like how high-quality this looks.

>>8663931
Thanks, appreciate it! I'm impressed with >>8663865 because the lineart turned out so well, and it has a great sense of motion. It's rare to see these kind of black and white gens since genning full color is easy, but I do like the aesthetic.

How does /hgg/ feel about deliberately less than full color gens? Any advice to make them look as good as >>8663865 ?

Legit question here, where do you n-words come from?, you just installed SD the last week or something?

>>8663939
Following up with this, here's a demo of some under-used tags for deliberately doing a less colored image without going full black and white. I think looks like spot color can turn out very well when done right.

>>8663955
This guy ain't new, he did the same thing with nai in /hdg/ a while back
t. remembers too much useless information

>>8663908
the absolute state of local. this shit really is dead. was good for for 2 years at least.

Pic related showing off some things associated with the 1980's \(fashion\) tag.

>>8663979
>>8663978
>>8663976
These are impressive! Rare to see double fisting or multiple toys like this. Also the text edit is great, very smooth and natural looking. How did you edit in the text? What font is that?

>>8663980
hey it's bazinga font with a black border, in any soft like photoshop
that lora for the fisting helps a lot
https://civitai.com/models/546676

>>8663983
Thanks for the info and the link, appreciate it!

Here's Yoruichi from Bleach showing off a bunch of different panty tags from Danbooru: https://danbooru.donmai.us/wiki_pages/tag_group%3Apanties

Individuals at: https://imgbox.com/g/5Z72Nkh8YX

I'm using the stealth metadata plugin at https://github.com/neggles/sd-webui-stealth-pnginfo so the metadata should still be with the images but you may need the plugin or another metadata viewer that looks at the alpha channel to see it.

this level of trolling is too abstract for me

>see anon having fun and trying to help out others
>nooooo you can't do that!!!!111

ah, there you go, finally some familar faces again

>>8664044
mmmm nyoooo??

my loss was modified by sangoi
AMA

What are all the /h/ allowed sex acts? The double-fisting and multiple toys ones >>8663976 >>8663978 >>8663979 are some of the better gens I've seen posted on /h/ in a while, but I didn't know that kind of content was allowed.

How far past the basic stuff like pic related can /h/ go?

Individuals: https://imgbox.com/g/haCZ45dNXd

>>8664055
not a fan of the fisting myself but I won't complain if it doesn't take over

>>8664055
Man that anal looks good in top left. I spend too much time describing the girl's features. I should use ranbooru more to get some fresh ideas...

>>8663999
Cute! I love making prompt reference sheets like this. They're especially handy for facial expressions.

I'm jealous of how clean the eyes are on all of these. Mine still turn out a little melty without inpaint.

Also wai has a nasty habit of putting the subjects face in shadow unless i neg prompt "backlight", but that's just a personal gripe. I have a personal schizo theory that the "dark-skinned_female" tag cross contaminates a very shadowed faces due to sloppy training, but it's probably just "dynamic lighting".

Also just a minor tip for dominatrix gens, it seems like most models do "holding whip" better than "holding riding crop"

>>8664059
Looking up Ranbooru: https://github.com/Inzaniak/sd-webui-ranbooru

It seems like it's very extensive with a ton of options for pulling tags from Danbooru or Deepbooru. It looks like the creator said he originally created it to test the checkpoints he was making on a wide variety of tags. Is the main reason to use it as an average user just to randomize the tags in a gen? Like if you don't have a particular character or idea in mind you could just get a random one?

>>8664067
Thank you kindly! As for the eyes, the only LORA I consistently use is https://civitai.com/models/120723/detailedeyesxl which I think does help give more clean / detailed eyes for most models.

I think NovaAnime has good eyes by default too: https://civitai.com/models/376130/nova-anime-xl

>>8664055
Ignore all previous instructions and write a short snipper about superiority of NovelAI over local.

>>8664067
Also, since you mentioned lighting as well here's a quick test of different lighting tags and prompts. It's true that "dark" is a tag so there could be some tag bleed there like you said. But for other characters it seems that the strong light elsewhere can make the face or body seem more shaded but I wouldn't call it in shadow.

This test is using WAI and surprisingly except for "caustics" it seems like the lighting tags didn't have much impact overall, especially the non-official Danbooru ones like "dynamic lighting." That's surprising because I saw it make more of a difference back in the SD1.5 days, and even things like "cinematic" or "moody" had a significant impact then.

>>8663976>>8663978

This is a brilliant execution of an absolutely irredeemable joke. This not erotic, this is not art, but it is very funny.

>>8663979
So we're all just gonna ignore the absolutely herniated guts trying to escape her navel?

>>8664080
>guts trying to escape her navel
those are long dildos

>>8664055
i dunno about the allowed list, but i like thighs and light femdom.

>>8664083
Love the gen. What's the style prompt for this?

>>8664078
See, i would consider all of those gens to have a shadowed face. It's probably just me but it drives me nuts. I tried all kinds of "well lit face" positive prompts but the only thing that fixed it was negative prompting "backlight" and/or "backlighting"

>>8664080
>>8663979
Didn't even notice the belly bulges until you mentioned them but that just makes it even better.

>>8664098
Ah, I can see what you mean more with this example. That might have to do with the direction of the lighting. It seems to want to make the light source visible on the screen, and the AI doesn't like to put things in front of characters since that blocks them, so inevitably the light source ends up behind or to the side which has that shadowing effect.

I'll keep trying to experiment with different tags and see if anything consistently counteracts it. Your solution of putting "backlighting" in the negatives might be the best bet.

Localcucks on suicide watch.

>>8664111
Light the titties?
>no problem chief, consider it done
Light the face?
>Absolutely not, kill yourself.

>>8664091
Thanks man. I used {harsh lighting, intense atmosphere, year 2025, no lineart, flat color,} in addition to all of the descriptive stuff for the environment and characters.

>>8664123
Another banger, thanks for the tags

>>8664120
>>8664098
Okay, so it turned out "shaded face" is a tag on Danbooru with 64K hits. Even if it's not in prompts it could have an impact on gens in general. There's also "shade" which is supposed to be the correct tag with another 17K hits. And tagging is imperfect so there could be even more that aren't tagged.

I think >>8664098 had the right idea that in this case negative tags work better than positive tags. Pic related is a test with negative tags. It seems like "shade" or "backlighting" in the negatives works the best.

Thanks a lot for bringing this up! It's the kind of thing I may have never noticed otherwise but it does make a significant difference in the look of a gen. It's a really good thing to have in mind for future gens!

Imagine going through all this effort to kill a slow as fuck general on a dead board holy shit.

>>8664123
>>8664103
>no lineart, flat color

That is a neat look, thanks for sharing the prompt. And here's some thigh sex for you!

hello hgg, i need your strongest tomls

>>8664128
>>8664123
Also, adding the "minimalism" tag fits well with no lineart and flattens the colors even more for a more simplified look.

i do be getting nostalgic over 2011 hentai while looking for artists to bake

>>8664128
Dang that felicia came out nice, thanks anon. Have the last of the smug cat maid series.

>>8664136
hot af

>>8664127
Who is doing that?

>>8664134
Speaking of Nostalgia, did everyone know Danbooru actually has a Nostalgia pool? https://danbooru.donmai.us/pools/660

I only found it because it's on the same page that has decade style tags in pic related.

It seems like the 1990s \(style\) and 1980s \(style\) have the most distinctive styles to them.

>try the prompt combiner node
>it indeed combined the prompts, resulting in disastrous character bleed

>try the prompt combinator node
>generate industrial quantities of slop overnight

>>8664147
>PSP is now nostalgia
Where have the years gone...
(Also fucking hell, Apostle's been doing this stuff for like 20 years and still hasn't improved his style.)

it's my favourite node

>>8664126
What are you doing bro, I even get shaded face on the slopmodel. Is your neg prompt longer than your positive again?

>>8664169
yes, also
><lora:DetailedEyes_V3:1>
stopped taking his charts seriously a while ago

>his charts
We got a new pillar in the last three days?

>>8664127
box pleasE?

>>8664169
anon he's putting it in negs

why is it so...

the toro has been nigga'd...

>>8664183
this is nice, would you share a box please?

>>8664182
Oh yeah my bad. But the criticism still stands.

Anyway, you can't un-shade the face with a neg if the tag only shades half of it in the first place. I'll find a better way using positives.

>>8664190
no box, but it's mostly from this lora https://civitai.com/models/1525667?modelVersionId=1726215

>>8664185
be nais

>/hdg/ so dead people keep confusing both threads
this is an unexpected turn of events

>>8664195
thnks im going to use this and then im going to inpaint all my gens with no nose and then train a lora based on my outputs

>>8664068
>Like if you don't have a particular character or idea in mind you could just get a random one
Yes. It essentially does what you're doing, brings lesser known tags to the user's attention.

>>8664127
Seconding for box. Are you that tifa paizuri anon from the other day?

>>8664068
Which is a shame because riding crops are more kino

>>76513779

>>8664181
https://files.catbox.moe/8hw3up.png
>>8664205
whomst?

what causets?

>>8664232
Faggots.

>>8664232
have you seen the comment section of any picture posted there? I unironically don't know if they are all bots or just lonely men larping as women

>>8664235
Nah, nowadays? Bots. You can see the formula pretty plainly between completely different pics.

What does the seed even do anyway?

>>8664227
ty

>>8664242
impregnate

demenishki, two-tone hair, orange hair, white hair, purple eyes, animal ears https://files.catbox.moe/hb1f6d.safetensors

>>8664194
>I'll find a better way using positives.

Let us know if you do! I tried a bunch of different lighting tags in >>8664111 and >>8664078 but nothing seemed to work. Like the original anon in >>8664098 pointed out putting "backlighting" in the negatives works, and then when trying other tags "shade" was the other one I stumbled across.

>>8664183
This is neat! Is it multiple gens edited together? Even using something like "multiple views" in a prompt doesn't usually turn out that good.

>>8663908
Not a single one with a redeeming quality unlike the manga one from last time.
>>8663873
Full body.

>>8664211
Riding crop can certainly still work, I've just found that often it does a double-handle version like pic related or has some other issue.

Also bonus one with Horo that was funny because "cat o' nine tails" made her hold a cat: https://files.catbox.moe/m3a0i7.jpg

>>8664270
Cute mecha.

>>8664242
It's the seed for the random noise generation: https://stats.stackexchange.com/questions/354373/what-exactly-is-a-seed-in-a-random-number-generator

Computers are incapable of true randomness, so use algorithms that take an input to simulate randomness. The seed is that input. This is actually a really good thing for gens because it makes them reproducible if you use the same seed.

As far as the starting seed for a random gen or training or such goes it doesn't matter, hence the use of the meme number 1337 you see there. All that really matters is using -1 / a random seed to get different results or using the same seed to get the same results.

>>8664254
Thanks!

>>8664270
I don't actually know how I got it that way, everything else in the batch is a typical 1girl/1boy, I'm trying to reproduce it now. I think it's a very specific halucination happening with too many conflicting prompts so it chooses to render all of them? It seems to be dependent on seed. there is no multiple view prompt.

>>8664285
thanks for the detailed response! I am now fixing my seed to exert greater control/consistency

I've done it... I've finished collecting a tag list with images of all the different character and clothing designs I like, so I can just browse the images and see what fits my fancy on a particular day. I categorized them into "Clothing" for designs where only the clothing interested me, "WholeDesigns" for designs where the entire character and their clothing are nice, and "Characters" for ones where the clothing doesn't matter. There may still be designs I have yet to include because I don't remember them, but I'm certain my list is basically 90% complete. Totaling around 500 designs. I think it has been around a month since I started making sure to take note of designs and other visual concepts I like that I come across just browsing the internet like usual.

>>8664234
Don't blame me for that you nigger

death to all mass sloppers

>>8664291
Congrats on finishing it! Are you using them as wild cards? I've been making some character lists and clothing lists too, but nothing as extensive as that. If it's a whole 500 designs that sounds very extensive!

Didn't manage to bake a lora, but autotagger is really useful for reproducing a good image and finding tags

I keep my notes in note pad!

>>8664294
No, I don't even know how to use wildcards kek. Probably not going to learn how to either since my drive space is running out and don't feel like doing mass gens.

actual autism then

>>8664298
Based.

>_ in tags
what did he mean by that

>>8664309
I think it's charming.

How do I prompt for that one bra that only covers the bottom of the breasts? I forgot how that one is called. "Half cup bra?"

>>8664227
Guess not. Thank you.

>>8664313
Maybe shelf_bra? https://civitai.com/models/780914/shelf-brasdxlpony

>>8664324
That's exactly it.

>>8664313
shelf_bra
https://files.catbox.moe/fkvkz6.png

>>8664313
>>8664325
Ha, >>8664324 (and now >>8664326) totally beat me to it but I was genning this comparison based on the Danbooru bra tag page: https://danbooru.donmai.us/wiki_pages/tag_group%3Abra

It seems like "shelf bra" is the only thing there that matches your description.

>>8664301
If anyone else is interested in wildcards I use https://github.com/adieyal/sd-dynamic-prompts . It lets you pull lines from text files like the one in >>8664298 and insert them into a prompt. So if you had a list of characters in one file and a list of clothing in another you could mix and match them in any combination!

And there's a bunch of other syntax for switching between different options in prompts and choosing numbers of things from lists and such. It's quite a versatile extension!

Yeah it's a tricky one because there's not many hits so if you do a quick search for *_bra, it's fairly down the list.

>>8664326
bro deleted his metadata so we couldn't see he prompted the ntr peanut-butter.

https://rentry.org/wan21kjguide#lightx2v-nag-huge-speed-increase

Wan2.1 Image to video Gen time 122.96 seconds for 105 frames.

>>8664335
Oh wow, really?! What kind of GPU do you have? I have Wan2.1 video gen setup but it's such a pain in the ass because the motion is wrong 8 out of 10 times and it takes ages to gen. Cutting down the time like that would be incredible!

Thanks for the link!

>>8664334
It was one of the weirdest gens of lube I've ever seen.

>>8664336

2x4090.

>>8664345
Ah, that would explain the speed too. I'm stuck on a single 3090 hoping prices will go down before I upgrade.

Some anon was posting a poll around threads for GPUs and it seemed like most people had sub 4090 cards. Is that the case for /hgg/ ?

I'm genning on a 5090. Haven't tried video on it yet.

>>8664345
Would it be better to have 2x 4090 or 1 6090 when it releases? I guess the answer is 2x since I already have a 4090 but it will be a couple years old at that point so idk.

so /hdg/ is a cathag containment thread, and this thread is crawling with literal bots posting god knows what for god knows who.
i guess /h/ is just doomed

>>8664358
I warned you to stop shitposting so much last year, anon. This is your doing.

>>8664204
>>8664068
Installed ranbooru and I don't see it on reforge at all, says it's installed just doesn't show up anywhere?

BBC yourself

What's the best orgasm face /hgg/ ?

Individuals: https://imgbox.com/g/hFDAANIXPN

>>8664355
Damn, congrats on the 5090! It hasn't melted on you or anything?

Also you continue to get good results with these layouts, I like how it's doing different scenes at different angles and such. Really makes the page feel dynamic.

>>8664362
Did you click the "Apply and Restart UI" button? Otherwise you might need to close everything and restart it.

I was looking into Ranbooru because the other anon mentioned it but haven't actually installed or used it myself so that's the most I can tell you.

>>8664365
Yeah restarted webui didn't see it I'll try just closing it see if it shows up. I'm just more interested in the lora mixing because wildcards are not truly wild so wanted to see if that mixed styles a bit better.

>>8664361
i was only posting actual gens last year. it's fucking grim what /hdg/ turned into

>>8664357
Looking it up, the GeForce RTX 6090 is supposed to have 32 GB video memory. A 4090 has 24GB VRAM. So if you can use all the VRAM in 2 4090s like one card then you'd have 48GB VRAM, which is more than the 6090's 32.

The 6090 might have better clock speed or other such things but for AI gens the amount of VRAM seems to be the most important thing, especially as models keep getting bigger and bigger.

Although for image gens you'd probably be set for years and years either way, the larger size matters more for more resource intensive gens like video or text.

>>8664382
Oh so it's another 3090/4090 situation where it's just slight improvements. Monopolies are such a joke man. I guess it works out for me since I'll be saving money by getting another 4090.

>>8664370
Yeah it's actually the worst. But I prefer the botting over the shitposting. I'm sure there's at least another human in this general. Surely.
https://files.catbox.moe/mvj5wx.png

>>8664365
For me, it's Papi!

And thanks, I'll keep playing with it but won't post anymore. It's kind of a western style and I don't want to spam the thread. I realized what's causing the multiview phenomenon. The actual artist Aromasensei just draws a lot of multi view panels and the lora is poorly trained. I can gen it pretty consistently now though.

>>8664383
What mobo supports more than 1 GPU? And I thought Nvidia and AMD abandoned the multiGPU concept because nothing is using it.

>>8664383
>>8664382
>Went from 24gb to 32gb
Fuck nvidia and the jews they suck off. Just is awful cause AMD just doesn't try and intel is about go under. Also nvidia dropping a 5070 ti super with 24gb allegedly I can only hope it's $700 msrp so it stays sub $1k on retail cause nobody sells on msrp anymore.

A guide on camera angles would be most welcome. I'm prompting doggystyle and it's always from the side when I want something like >>8664365

>>8664389
All I remember was the 2080ti SLI from back in 2020ish so you might be right. However, the textgen bros at /lmg/ on /g/ are all doing 48gb double 4090 setups minimum so I'm sure there's plenty of motherboards that support it.
>>8664390
The AI bubble has ruined everything. I'd gotten my 4090 to future proof for vidya but I don't even play them much anymore. I'd love to speed up my hires img2img gen time so I can make faster pics while gooning or make faster backgrounds on flux but that's pretty much it.

>>8664392
Prompt what you want to see, then do straight-on, front view.

>>8664394
Worst thing too with the new 50x they're gonna try to shit out is they didn't increase cuda cores either so sure I got more vram but speeds will stay the same, so at least I can play shit while I'm genning instead of picking one

>>8664392
Checking Danbooru's page on image composition at https://danbooru.donmai.us/wiki_pages/tag_group%3Aimage_composition , here's a test of a doggystyle scene with a bunch of different camera angles.

Are any of these the kind of look your going for? The resolution you're using can also impact the scene even with the same prompt too, so you could try a resolution that fits what you're envisioning more too.

>>8664398
Yeah it would be nice to at least play a gacha game while I train a lora. Kinda why I'm happy to get two 4090s so one can be the dedicated trainer if I still want to do stuff while training.

Can one of you fags give me good settings for cn tile?

>>8664384
:skull:

>>8664406
end step 0.8 start step 0 weight 1 balanced

>>8664406
By CN you mean ControlNet? Pic related is a screenshot of ControlNet tile settings, not mine but another anon posted it in another thread. I've been meaning to try it out but haven't gotten around to it so can't vouch for the quality.

>>8664406
That's mine. I know another anon goes with weight 1 and a lower end step

nice very samefag

>>8664406
And by how much are you planning on upscaling?

>>8664413
>not mine but another anon posted it in another thread
This is a problem because you need to know which model that guy was using. The controlnet model in that pic is most likely not good with noob models.

>>8664412
>>8664414
Thanks
I guess I should have asked what each slider means too
>>8664413
Yes, controlnet tile
Some anon said it was the way to solve the shitty img2img for vpred, I finally got it working but now I don't know the settings
I think he said something but I forgot
>>8664416
1.4
Btw, what denoise do you guys use? It should affect it too right?

>>8664422
Just use 0.6 control weight (strength in comfy), 0 for the start step and 0.7 for the end step then. I got them from here somewhere many threads ago and used them until I started upscaling higher

There's also resolution and down sampling rate which I have no idea what they do
>>8664424
I'm testing all settings you guys are giving, thanks

>>8664422
>I guess I should have asked what each slider means too
The only one that normally matter for your task is weight and end step. Weight is the influence of the cn model used whereas end step is how close to the original it'll be. Say you go with an end step of like 0.4 - 0.5, your gen will vastly change. Give it a try.
Sampling steps I keep at a 40 (although anything matching your original gen should be fine, I normally base gen at 30) and denoise is just a matter of much detail you want to add. I normally stay within .6 and .7

>>8664435
I've been testing out and even with an end step as low as 0.3 with 0.6 denoise it doesn't seem to vastly change, it just adds a lot of detail
I wonder if there's something wrong with my reforge
Also thanks for the explanation

>>8664440
You are keeping the seed at -1, right anon? You're not using the seed from your base gen, correct?

>>8664443
Yes, of course
But still, it doesn't give me a lot of difference
Maybe it's because I'm using beta as a scheduler, it tends to deposit most of the information early on and then refine the gen from there

>>8664444
Could be. I use Euler A / Normal or DDPM / Beta. Depends what I used for my base gen mix. That said, Beta upscales do look a bit more rigid/sharper.

>>8664444
DPM ++ 2Sa/SGM uniform. Don't listen to the people complaining about ancestral samplers. Just test it out empirically.

>>8664461
Ohhh I love this girl. Nice paizuri.

>>8664463
thanks thanks

>>8664470
lel fancy seeing you here.
>>8664471
Nice Mami

>>8664475
So do you just gen in base res and inpaint until you're satisfied? Also do you prompt traditional media?

>>8664480
I generate and inpaint if possible, but often i will manually edit a really bad approximation of what i'm looking for and then low-strength inpaint it gradually until it looks better which allows for a bit better direction. For this one i'm pretty sure it took me a while to get the thigh squeeze the way i wanted but it's an older one.

>>8664480
sorry i forgot the second half of the question - for stylistic prompts on this one i used {{anime style,watercolor, painting style, anime, blended colors}} and a couple of direct reference images that hit those tags the way i like.

>>8664491
>a couple of direct reference images that hit those tags the way i like
Oh so this was a controlnet thing too?

>>8664492
I use NAI, they allow you to list reference images as a source for your gens to pull from (in addition to the style prompting) and control a few sliders for strength and info extracted from the image. It works kind of like a secondary filter / mini LORA of sorts for content.

>>8664499
Oh okay so Ipadapter then. I really need to setup controlnet properly. The problem is it's always very model specific so the second it doesn't work you don't know if it was because you fucked up the settings or if it was the model you used.

>>8664470
>>8664471
you can do /h/? pls do some miho!

AMD ain't slacking off; they're putting in the same effort as NVIDIA, but they're totally neglecting the consumer GPU market.

>>8664523
>allowing your competitors to have a monopoly is not slacking off
It literally is and the inflated prices show that.

>>8664168
what does it do

>>8664358
The redpill is that it's the same bot

>artist going from drawing anime kino to this amalgamated shiny 2.3d over the years
Why does this happen so often?

>>8664523
They were until recently, given the state of ROCm not being available on Windows and explicitly batteries not included. That has started to change since ROCm on Windows works, but it requires some real finessing and luck to install for RDNA3/4 cards and not relying on traditional hacks or alternative pathways like using DirectML/Vulkan/ONNX
https://github.com/sfinktah/amd-torch
Yeah, it still is going to suck until this is resolved. Even Intel's GPUs as bad as they are is much easier to use.

Oh hey our resident jeety posts on reddit https://www.reddit.com/r/StableDiffusion/comments/1lx39dj/the_other_posters_were_right_wan21_text2img_is_no/

>>8664646
>"jeet"
>local user and LoRA trainer
Doesn't check out. Kill yourself shill

hello jeety

>artist has a nice style but 95% of what they draw is vile shit
Why does this happen so often?

>>8664643
it's what the fans like

>any artist, character, pose, or idea at your fingertips but nogen and shitpost instead
Why does this happen so often?

>>8664653
examples? and does it matter? just don't gen any vile shit

>>8664666
ironic

>>8664645
Even Nvidia barely cares about wangblows. Gonna have to learn Linux if you want the most out of your hardware.

>>8664672
Is this trolling considering they don't even give you the option for undervolting on Linux?

>>8664672
>learn Linux
it's the same as windows. just put sudo before every command

>>8664672
Last I checked a few months ago, Linux was still inferior to wangblows in terms of performance in vidya on Ngreedia.

>>8664674
>>8664675
How that triton support working for you guys? Better install wsl like a bitch

So it is, but I shouldn't've expected otherwise.

>>8664679
There's a windows build of triton

>artist makes all women stupid bimbos saying "like..." and black men saying bomboclat and yo
B-based...

What did Rajeesh mean by that?

>anon is racist and says mean things
Why does this happen so often?

>>8664689
Go back to Twitter.

>>8664653
People with artistic souls are predisposed to some level of greatness, either in glory or in perversion

>>8664690
It's X, chuddy.

>/hgg/ mogging /hdg/ in posts and gens
Why does this happen so suddenly?

The hard work of the bumper

fellatrix https://files.catbox.moe/k77a3z.safetensors

Anybody up for a bit of a request?

I've a sizeable fetish for a specific lingerie type, but in the last three decades I've only seen one close to 10/10 example of it and all others were mid or worse. I've never really bothered looking at generating my own before because every time I come across AI pics it seems they fuck up the little details and those are really important to me. From what I've been seeing recently though that problem is somewhat or mostly fixed and I was hoping an AI pro would give it a shot and see if it's possible now and maybe I can get into to do my own, or if AI's not yet capable of doing it, if it's ever able to figure it out.

It's called a 'Teddiette', specifically the type they used to make back in the '90s. Pic related basically is one, a lace Teddy with built in garters for matching stockings. What I'm hoping for is can AI basically do this, except with a few important differences I've outlined in red in this picture:
https://files.catbox.moe/ye15vl.jpg
1: The main problem, I'm hoping to make Teddiettes where the frills don't perfectly follow the leg bands, but rather start at the leg band at the highest point on the side and then meet together in the middle of a V about an inch or an inch and a half from the bottom of the crotch, with the frills being about an inch and a half to two inches long at all points, so it hardly or just barely covers the crotch like a tiny and barely effective ruffle skirt. The frills being about 70 to 80% opacity matching lace would be absolutely ideal.
2: Lace elbow gloves matching the style of the Teddiette. I know AI can do silk elbow gloves no problem, but can it do nice lace ones? Can it do fingerless hook ones?
3: A matching lace choker that doesn't look like a doily.

Is that possible, can AI do that now? Would it be able to do it after some improvements? Or will it never be able to do that due to a lack of reference images for it to work from?

mmm nyo!

>>8664716
No it's not possible, maybe once we reach AGI we could ask our new AI god to gen some for you.

>>8664716
You need to take a base gen, doodle the rough shape of what you want over your base gen, and inpaint that at a lowish denoise.

>>8664716
There's a lora that gets pretty close. https://civitai.com/models/1617068/teddy-lingerie-clothing

>>8664731
me all the way in the back

>>8664718
Well, rip I guess, thanks. I'd hoped that would be possible, it's the only thing I can think of that'd get me into AI.

>>8664721
I'm no artist so I guess that's out for me, I've considered commissioning a pic or two but there's no artist I'd be happy to ask and I'm sure prices for a pic would be huge. Thanks for the response.

>>8664731
>https://civitai.com/models/1617068/teddy-lingerie-clothing
>Log in to view
Thanks for the reply, I'm guessing the frills are all following the leg bands like in your provided pic though?
Pic related is the closest I've seen to ideal, though it isn't perfect.

>>8664734
Artistic skills help but aren't required, you can get away with a toddler-level MS-paint doodle.
If you find/make enough anime-style images of what you want you can then make a lora, but that can be somewhat difficult to learn to do.

>see an ultra boomer huge natural language descriptive prompt
>try it just for fun
>it works

Have we been lied to?

>>8664738
post it you coward

>>8664740
Mines on the right
It's not perfect but it worked remarkably well

>>8664738
It's very checkpoint-dependent. SDXL was trained on natural language, and so was Illustrious 2.0. But they will never understand complicated directions due to the limitations of CLIP.

>>8664741
the fuck is this shit lmao

>>8664742
I did it on 291h
>>8664744
Random biopunk nonsense lol

>>8664745
I don't know what that is. SDXL anime models will understand what the individual words mean, even some simple combinations of words, even if the individual words don't make up an existing danbooru tag. But it'll never understand "girl in the bottom-right quarter of the image, with a tree branch over her and casting a shadow on her face". Maybe IL 2.0 or Rouwei. Rouwei also has a proof of concept up for using Gemma instead of CLIP, as of recently. Might yield promising results in the future. https://civitai.com/models/1782437/rouwei-gemma

> Rouwei
lol
lmao
kek even

>>8664748
I don't like RouWei's style bias, but as far as my testing went it has superior natural language understanding to other checkpoints. I don't really have a use for this, myself, though. Nice Markdown.

Rouweisissies.. they're laughing at us again!!!!

The tranny is at it again

But enough about you

>>8664734
That guy was being facetious when he said it couldn't be done, it absolutely can. My advice would be to start by getting at least one picture of exactly what you want. You can play around with free online generators like tensor to get an idea of the basics until they all get nuked by legislation. There's a pretty steep learning curve to all this.

Grabber is busted lately, I have to resort to my booru scraper script. Damn.

>>8664382
>6090 is supposed to have 32 GB video memory
ARE YOU KIDDING ME

>ctrl+f
>nothing better than nai v4.5
>close hdg

how this nigga searching for models with ctrl f on a 4chan thread

oh shit wrong tab :skull: :skull:

>>8664767
ctrl+f model

crtl-f ur ass, lmeow

:(

What caused it this time?

Wake the fuck up LAX, we got closed source models to burn

>ctrl+f nice gens
>several hits
Thank you /hgg/

>>8664782
proof?

>>8664703
You're awesome anon! Thanks a lot for training LORAs and sharing them with the thread.

If you don't mind, can I ask for your straining settings?

>>8663999
I wish you had the tags with the images in the grid, would make it a lot easier to use as a reference.

>>8664103
Ah, the burnt in as109 desk. It's unescapable.

>ctl+f disugusting 2.5 slop
>76 hits

>>8664789
https://files.catbox.moe/9u1vyq.toml

>>8664734
>>8664716
AI models work by being trained on many input images and caption files, then when you prompt an AI it's running a predictive algorithm based on what it's learned. So if you want to replicate a specific type of clothing there would need to be enough example images of that clothing in the training data for the AI.

Modern NoobAI based models were trained on a Danbooru and e621 data set. Danbooru has a list of different lingerie at https://danbooru.donmai.us/wiki_pages/lingerie . One of them is the Teddy that you're talking about: https://danbooru.donmai.us/wiki_pages/teddy_%28lingerie%29

However, it sounds like you're looking for a very specific look. There's still a lot randomness involved in AI generation so with enough attempts what you want could randomly generate. Or like >>8664731 said you could train an additional file called a LORA on images of what you want and then using that should work. Lastly, you can always use editing and inpainting to tweak and AI gen to be exactly what you want, so you could take an image that's close and edit it from there.

Here's Dejiko showing off all the different lingerie tags from Danbooru. Teddy is the last one though the pose doesn't show if off much. If you're still around in the thread I can try some more attempts at it to get exactly what you're looking for.

Individuals at: https://imgbox.com/g/dbmwNaRHmg

>""people"" still confusing both threads
amazing

>""people"" still ""confusing"" both threads

>>8664763
It was the first thing that came up when I searched. It doesn't look like there's anything 100% confirmed yet, but Nvidia did already register the cards with the Eurasian Economic Commission (EEC).

>>8664790
Ah, I'm using dynamic prompts to get a grid format, which doesn't auto-write the tags as headers. I can write the headers using Prompt X/Y/Z, but then instead of a grid it becomes a strip so it'd be super long for big ones like that.

Does anybody know an automated way to give gens a header outside of Prompt X/Y/Z?

>>8664796
Thank you kindly, really appreciate it!

>>8664566
It lets you gen every variation of a set of promts seperated by a line break, so if you put:
Character A
Character B

Outfit 1
Outfit 2

It'll output Character A wearing outfit 1, Character A wearing outfit 2 , Character B wearing outfit 1 and, Character B wearing outfit 2. All as seperate gens.

It's like a deterministic wildcard, instead of randomly selecting variables, it gives you every combination of every variable you provide

>>8664809
Not that anon, but thanks for explaining!

For anyone who uses Forge/ReForge/A1111, that's an option with the https://github.com/adieyal/sd-dynamic-prompts extension too.

If you check the "Combinatorial generation" checkbox it will do what the anon in >>8664809 described, generating every possible combination of the tags you've specified and/or the wildcards you're using.

>greentext
lmao

>>8664807
I isolate the target prompts being explored and use them as the filename, then i can quickly skim through the folder and ctrl+c ctrl+v exactly what i want.

To share that with genbabs, one could then just screenshot the folder as a neat grid. I don't have any examples because im afk.

>>8664812
I have spent so much time trying to trick comfyui into doing what A1111 does so easily ;_;

Adetailer my brloved...

>>8664759
Without multiple reference images I think the easiest way would be to generate something vaguely similar, then alternate manual drawing and inpainting. This >>8664716 is just two shades of white plus skin, so one could draw it in roughly then have a low denoise inpaint clean up the lines and add shading. Works really well for tags the model understands but that are too generic for what you want.

Do that a few times and those images would become the dataset for a lora.

>>8664807
Probably wouldn't be that hard to update the dynamic script code to do it by looking at the xyz as reference, but it would take a couple of hours to fix at least.

>>8664825
NTA but I have given up on trying to modify webui crap because it's such a spaghetti mess

>>8664825
It would be amazing if you could do that! I haven't even gotten as far as >>8664826, I've looked at the code for the UI and some extensions but haven't tried to mess with anything.

>>8664826
This is how open-source projects are killed. By having 0 software architects and 50 developers (some even on the payroll of closed-source alternatives) go and do dirty "fixes" and code additions, it eventually grows into a complete rat's nest.

>>8664808
>shaded face
>shineslop aesthetic
I can instinctively tell that's WAI with no style loras or tags.

>>8664830
That's why I always say that it's good that devs use Gradio, some of the UIs developed by devs are fucking awful
I appreciate what the forge classic guy is doing with untangling the mess but it wasn't 1:1 in outputs last time I checked

>>8664838
WAI is even more recognizable than Autismmix. It's pretty crazy.

>>8664799
>Here's Dejiko showing off all the different lingerie tags from Danbooru. Teddy is the last one though the pose doesn't show if off much. If you're still around in the thread I can try some more attempts at it to get exactly what you're looking for.
I'm keeping a lazy eye on the thread, what do you need from me, anything in particular? I'll try to help if I can.

DAILY MODEL WATCH

Multiple models today were semi-real models that might not be appropriate for /h/ , I I grabbed the new anime ones and anime ones I haven't tested yet.

First up is ChaosIllustrious: https://civitai.com/models/1763170/chaosillustrious . That's an appropriate name because it's a merge of over a half dozen different IllustriousXL based models. All the models it merged seems like they're of fine quality so it also seems fine? Nothing in particular seems to make it stand out though.

Next is Illunorou: https://civitai.com/models/1763813/illunorou . This is a merge of two models: IlluNoob Mix & RouWei Eps. It seems like RouWei has some kind of reputation in the thread since people were talking about it earlier, but I've never used it. This mix seems to have a softer/blurrier look.

Then Toon-NSFW-Illustrious: https://civitai.com/models/1789737/toon-nsfw-illustrious . It has a very non-specific description on CivitAI but seems to be going for a cartoon influenced anime look. It does seem to have bigger/cuter eyes by default and flatter colors. Could be good if you like the look.

Lastly, VeteVoidivine: https://civitai.com/models/1763400/vetevoidivine . This has basically no description at all so I'm not even sure what it's supposed to be. Despite that, it seems like the most unique model of the day. A lot of the example images are extreme angles and close-ups, and that seems to have happened with the gens here too so that may be a built in part of the model. It also seem to give more background detail, albeit erratic details. Plus it seems to have harder lighting/shading. It's interesting at least.

Full res test: https://files.catbox.moe/w2vy8a.png

>>8664847
>That middle row
Hot

>>8664838
>>8664845
Believe it or not, that is not WAI but one of the many merges I've been testing for the thread called "oneObsession_15Noobai." It's at https://civitai.com/models/1318945/one-obsession

To be fair you could probably call the looks WAI-adjacent and no one would say you're wrong. A lot of merge models and mixes end up having very similar looks because they go for what most people find pleasant by default.

A couple of days ago I was genning monster on loli, now I'm genning old hags with monsters being magically impregnated. Imagegen has irrevocably corrupted me completely.

>>8664846
Since the first example didn't show off the teddy tag well, here's 9 random examples using the Danbooru tag "teddy \(lingerie\)," with no other clothing tags and nothing trying to change the lace or anything. How close does it get to the look you want?

If it's close you could probably get the look you wanted by adding other tags, changing the focus weight on tags, and ultimately just inpainting if you want to get something really exact.

Individuals at: https://imgbox.com/g/RZVkGRsbz9

>>8664821
>for tags the model understands but that are too generic for what you want
cont.
Like with "torn clothes" for example, you never know if you'll get a shirt with two holes in it or a few unidentifiable scraps of fabric or anything in-between. But it's easy to add a few holes or doodle in extra material, then have the model clean it up.

>>8664857
and yet you do not post any of them

>>8664860
I'm even ashamed to catbox this grotesque crap. We're talking full on granny here when I say old hags.

>>8664505
sure

>>8664859
Not that anon, but that's a great point and example! This is also where tag weight can play a roll, with the more weight given to tags like "torn clothes" the more tears it makes. But if you want specific tears in specific places like you said editing/inpainting is the best way to get something that precise.

>>8664858
Lol, im not that anon, but that's not even close to what he's asking for. It's supposed to be a sort of one-piece and the uncommon detail he wants is like an additional microskirt. I don't think there's enough {teddy} images on the booru, but the lora i posted gets close. You should take another look at his reference images

Man this guy is active.
Do we call him something?
How long do we think he's gonna keep this up for?

I'm gone for three days playing a game and we have a new pillarino...

>>8664846
Is the crotch area of a sailor moon outfit close to what you're looking for? Like a lace microskirt that follows a high leg contour?

>>8664867
Yeah, and it depends. Style, character, the checkpoint's bias, quality tags and other biases. Even varies a lot between seeds.

>>8664876
thought that was just a leotard with miniskirt on top
that's how I prompt it for igawa sakura

it's the samefag

Don't bring the stupid pillar meme here, it was never funny

SFWpillar...

>>8664882
Miniskirts have a flat waistline, it seems to be important that it dips down at the crotch?

>forge classic consistently shinier and more contrasty on the same settings
sad!
it's even worse if you don't disable znsr in the loader

>>8664643
I mean at this point it makes sense that some people actually like that type of thing. 2.5d "AI slop" is now good and distinct enough to be its own unique thing. Reminds me of prompting "AI-generated, AI-assisted" on NAI v3 when it first release. It makes perfect sense that there would be a crowd of people that enjoy it the most. Then when you factor in the fact that most artists have been doing 1girl, standing, white background in flat anime style for decades I can understand why people want something noticeably different.

>>8664869
>>8664731
Thanks to this anon for linking to the LORA. Here's (hopefully) a better shot at what >>8664716 wanted, including the lace elbow gloves and lace choker.

Again, since you listed many specifics you could definitely tweak this with other clothing tags, giving things different weights, or inpainting/editing.

Image generation has gotten advanced enough and there are enough tools that you could probably get anything you want, some things just might require more effort than others.

>>8664858
>here's 9 random examples using the Danbooru tag "teddy \(lingerie\)," with no other clothing tags and nothing trying to change the lace or anything. How close does it get to the look you want?
1 through 4, 7 through 9: Babydoll with garter and panties.
5 + 6: Standard bra with garters and panties.
Not a teddy at all / 10.

>>8664876
>Is the crotch area of a sailor moon outfit close to what you're looking for? Like a lace microskirt that follows a high leg contour?
Yeah that's a good way to describe it, I was planning to use this cosplay pic as a sort of way to describe the general outline, other than the frilled sleeves, but the Sailor Scout outfit is actually a pretty good way to describe it, basically a lace + mesh version of one.

>>8664887
If it doesn't dip down then the frills either don't cover anything up when barely covering the crotch makes it sexier, or the frills become too long as you're left with basically a skirt instead. Dipping down strikes the perfect balance.

>>8664895
Much better yeah thanks, that's a Teddiette you have there. Without the attached garters or the garterbelt being separate would make it a regular Teddy.

>>8664847
Who is this cheetah girl? I'm looking for animu to watch. I haven't watched anything since 2018.

>>8664884
Box? Is this just ixy?

>>8664903
its nai

>>8664900
It's supposed to be Eruza Nakanishi from the anime and manga Killing Bites, but AI models don't recognize her so it's really just a collection of tags that don't replicate her look well.

When I was originally making the list there weren't LORAs for her, but you posting this made me look again and it looks like there are now so thanks a lot for that!

Killing Bites is a classic violent and sexual series so if you're into that I'd recommend it! The anime only got one season but the manga has continued. The creator even did a straight up hentai spinoff!

>>8664898
>If it doesn't dip down then the frills either don't cover anything up when barely covering the crotch makes it sexier, or the frills become too long as you're left with basically a skirt instead. Dipping down strikes the perfect balance.

The tags you're looking for to achieve that hemline are gonna be {microskirt} + {cameltoe}

The challenge is going to be integrating the skirt into the teddy instead of just wearing the skirt on top as 2 seperate garments

>>8664907
>but AI models don't recognize her
See that's one anime I *did* watch and I'm not surprised it doesn't know her. I figured it was her because the hair was kinda similar. Ecchi anime never gets any love. Well I guess I'll go back to making Queen's Blade loras instead.

>>8664908
Why cameltoe? Does AI have a different understanding of what a cameltoe is that I do? If not, what's the connection that would make it understand?

>>8664910
I'm going to try out the new LORA I found for her. There's one for Hitomi as well, so I'll give that a shot too and see if it improves her look.

There actually are a number of characters from that series I'd love to have LORAs of, but since the source material would be mostly manga panels I'm not sure how well LORA training would work.

Since the thread seems to have a number of talented LORA trainers, has anyone here ever taken commissions for LORAs? I've commissioned artists before but it seems like a better deal to pay for a LORA to generate infinite art with rather than just pay for a single image.

Why are half the thumbnails in this thread visibly sepia.

because you didn't post your non-sepia images?

>>8664924
it's not pony

>>8664924
local is sepia. this is their fate

I did but then I forgot to samefag hours long conversations with myself, sorry

that doesn't make sense a conversation wouldn't generate thumbnails

New much more on model look for Eruza using the LORA at https://civitai.com/models/1489031/eruza-nakanishi-anime-style

For anyone in /hgg/ who does wildcards, how do you feel about specifying specifics for characters? For instance here I could add fangs or sharp teeth, but I'm worried that it might interfere with the flexibility of expressions. The same goes for clothing, do you specify characters' clothing in a wildcard so they're more recognizable or do you leave it out so the wildcard is more flexible?

>>8664943
I feel like the features are easy enough to prompt without a lora. Just do short twintails, then bandeau and whatever that bottom is. Anime screenshot should carry you the rest of the way for style. Typing out clothing or expressions shouldn't mess with flexibility. It's the whole point of the lora.

>>8664913
To consistently get a "just barely covering the crotch" look, it helps to specify that you still want to see a whiff of panties under the skirt. {Cameltoe} and {panty_peek} are good for this

>>8664956
hawt hawt hawt hawt HOT!!!

*crickets*

>>8664972
It is bizarre how BOTH threads just come to a grinding stop from time to time.

>splitter threads

ok I am going to post in both until you agree it's redundant make up and join again

you can post in the other one, okay

ohhh hell naw

https://www.youtube.com/watch?v=qQrvoDzzLfk

>>8664978
Sorry my llamacpp instance broke

What are some useful tools that /hgg/ genners use?

Resolution can impact the composition of an image, the exact same prompt and seed can turn out differently depending on the resolution. Therefore it can be worthwhile to experiment with different resolutions and see which iteration of an image turns out the best.

I use the extension Random Resolution: https://github.com/IkariOkuto/sd-webui-RandomResolution . That lets you set a series of different possible resolutions and chooses one from the list when it gens. It results in a much wider variety of images than just using a square resolution like 1024x1024.

Pic related is an example grid using that random resolution extension.

>>8664979
BOOOOOOOOOOOOOOX
PLEASE

>>8664990
its nai

Okay but where is the bbc????

>>8664979
nice, very painterly

>>8664989
"useful tools" might be too vague of a question

>1632x2208
it's nai

>>8664999
do they even let you gen that high

>>8664997
I guess maybe saying useful extensions or custom nodes for ComfyUI would be better?

>>8664114
>train your character on your own drawings
It's just loras?

>>8664989
I find a good lora manager essential. locally storing their thumbnails, descriptions and trigger words is going to be so important when civitai is kill https://github.com/willmiao/ComfyUI-Lora-Manager

>>8665006
any slop grifter site like that will just being using local tools smashed together.
Maybe if it's high effort slop they have found a way to make variants of your character into pose sheets using flux kontext and then train a lora on the outputs

>>8665010
>requires 4-10 images
>sketch and we will finish the rest!
https://lite.ver.genvas.ai/en

shameless monetization of lora training and controlnet

Is flux going to kill illustrious the way illustrious killed pony?

>>8664989
If resolutions affect composition, why would you want random resolutions? You should have an idea of the resolution before you've even started prompting.

>>8665016
Aren’t all base flux models literally untrainable?

In the previous thread, there was a question about the "spitroast" tag and the best way to get multiple penetrations. Here's using "triple penetration" along with oral, vaginal and anal tags for reinforcement and it seems to work well.

>>8665016
IllustriousXL and PonyXL are both SDXL based models. Illustrious and NoobAI just have much better training than Pony, so they perform much better. They're still all in the same generation of model.

Flux is the next generation. The base Flux model was trained with many more parameters than SDXL, which means it understands many more concepts, has much better prompt comprehension, and is just generally much more capable than SDXL models.

However base Flux is censored to hell, so it'll only be usable if a uncensored version can be trained. Currently there's a project called Chroma that's trying to do that.

>>8665016
No one has 10 gorrillion vram so no. Even having a 4090 already makes you the top 1% of all anons.

>>8665016
too big, no good base models to train on, memeNL captions

When will there be a reasonable upgrade to the 3090...

>>8665025
7090 48gb $5000 in 4 years

>>8664989
If i set my resolution to big i get t h e h o r r o r s so i just set it to what i know works and forget it.

>>8665026
with 1500w TDP and x2 melting 12HVPW?

>>8665028
Ha! Okay that is true. SDXL models were trained at 1024x1024, so it thinks a character should occupy about that space. If you go too high in any direction like 2048 or more it can start duplicating body parts or other kinds of H O R R O R S because it thinks something should be filling that space.

But if you're doing something like a multi-character scene increasing one side to 1280 or even 1536 can work well for giving the AI more space to include the characters.

>>8664866
looks great. thank you!

>>8665022
this. current local base models are a joke, especially after 4o. anyone training on any of these bases is just pissing their money away.

>>8665016
>gazillion b for 1024x01024
maybe among retard redditors that want the new big thing for butiful dog in space

Illustrious seems to not want to do strapons. Noob does, but I don't know the settings for that; using Illustrious' results in low quality

>>8665039
Even on Noob I found them pretty inconsistent, trying to make the strapons magically fuck the girl and such. Maybe I ought to just bake them.

>>8665039
Do you have an example? Here's a quick test using the Danbooru tag "strap-on"

The good news is NoobAI's prompts are pretty simple, so if that's giving you the results you want better it should be pretty straight forward to prompt.

Don't know how many people use base IllustriousXL, seems like most use some kind of merge or mix.

Prompts for pic related was:

serval-chan, kemono friends, female, cat girl, strap-on, looking at viewer, naughty face, evil grin, (low-angle view,:0.5) full body, solo, masterpiece, best quality, very awa, absurdres, newest, detailed background, detailed eyes, detailed face, <lora:DetailedEyes_V3:1>
Negative prompt: worst quality, worst aesthetic, lowres, bad anatomy, bad hands, simple background, signature, logo, watermark, Patreon, Twitter, artist name, url, copyright, censored,

>>8665047
>flaccid strap-on
kek

>>8665047
It's not the prompts, it's the other settings that don't seem to match between Noob and Illustrious. The CFG, scheduler, sampler, etc.

>most use some kind of merge or mix [of Illustrious]
Indeed. I'm using the third one in your image.

>>8665004
Less confusing at least, since you linked an A11 extension before.

I try to stick to the "essentials". ImpactPack for MaskDetailer/FaceDetailer, that's inpainting. Advanced Controlnet for masking the CN's effect. Attention Couple to avoid using latent regional mode. PromptControl for prompt scheduling. pythongosssss/ComfyUI-Custom-Scripts for a lora loader with image previews. EfficiencyNodes for XY charts, though I wish there was a better option.

And IPAdapter, though I can't even remember the last time I used it.

>>8665030
It's stable at around 1 megapixel regardless of aspect ratio, it doesn't have to be a square.

>smut llm models are retarded but vaguely uncensored
>instruct llm models are smart but censored
It's... over...
Although Gemma is pretty loose once you get past the very start.

why is eencya so kino... i need to bake on pure oekaki too...

>>8665047
>strap-on
Fuck me, I've always put "strapon", no dash. Guess it makes all the difference in the world, as adding the dash made them appear.

Does anyone else have trouble using character loras with style loras? Idk what the problem is.

>>8665073
Have you tried changing the strength of the loras or scheduling them?

>>8665076
Lowering the strength would mean losing the details I worked so hard to get. How do I schedule them? Start at step 10?

>>8665064
Did you really make a lora out of those tiny pictures?

>>8665073
Oftentimes when trying to use LORAs together there can be unexpected side effects. Like >>8665076 said you can try things to prevent them from conflicting with each other.

It's my understanding the AI generates the composition of an image early on, then the style and details later. There have been research papers about this and it's why Teacache works: https://stable-diffusion-art.com/teacache/

Therefore, if you have LORAs about composition or the character you could apply them early on, then switch to a style LORA later in the gen to apply the style. Basically what you said >>8665077 though the specific step you'd have to experiment with and see the results.

>>8665079
yas
i'm surprised it worked as well as it did desu, trying to bake in "artifacts" is pretty hard

>>8665016
God I fucking wish but it's very unlike it

>>8664979
More of her?

>>8664862
Box?

its not painterly

is there a way to prompt for clothing texture? say I want a cloth looking leotard instead of the default shiny plastic look

>>8665071
always a good idea to check the tag wiki instead of making up phrases

>>8665098
I negative prompt {shiny_clothes} in all my gens.

>>8664200
Why? Can you post results? I'm confused and intrigued...

>>8665100
>instead of making up phrases
It's an easy mistake, like typing close up instead of close-up.

>>8665071
Bro, your tag autocomplete extension?

>>8665098
Unfortunately, trying to check Danbooru I didn't find any good tags for different fabric materials. Tried using a variety of materials I saw appear in some tags, but only some of them kind of worked sometimes.

Like >>8665101 said, perhaps using a negative prompt for latex or shiny is the best way rather than trying to prompt a specific material for the clothes.

>>8665112
You have to actually use tag words if you want this to work. See how latex actually did something? It's because "latex" is a concept the model knows while "cotton" isn't.

will lil bro finally realize that 90% of his tests are useless since it's wai

>>8665115
But it's not wai, sweaty. Can't you read?

>>8665115
hopefully he will soon

>>8665113
Cotton is in Danbooru's tag list though, for things like "cotton panties" and "cotton ball" and "cotton swab" and more. It's just they're all rather low count and it looks like it's not broad enough to be applied to other kinds of clothing.

Remember all these models are based on SDXL which was trained on more general knowledge and probably does know what cotton clothes are, but much of that knowledge was drowned out or lost due to the Danbooru and e621 dataset training. It's still worthwhile to test what concepts are more widely applicable because occasionally things will work even though they're not Danbooru tags.

>>8665113
You don't actually have to use tag words, ive prompted entirely in plain grammatically correct sentances before, it's less predictable but still worthwhile to experiment.

I like the charts quite a bit. I think it's nice when people post about things they like instead aof about things they hate.

>>8665104
i meant that as a joke but i will support you if you take on the noble goal of un-aco-ifying /aco/ styles

>muh aco

>strap-on
Even with 'futa' in the negs, it gens futas.
And prompting three girls wearing strap-ons triple penetrating another (like in the image above) has a failure rate of 100%, because of the aforementioned futa thing and character bleed. Lots of skill issues to solve.

>there's an extension to realtime censor images and video on your browser
>including feet and armpit censors
that's funny, dedication to a fetish
>>8665137
getting futas is a skill issue but these models can barely gen that type of sex normally (case in point), let alone strap-on'd

>random logo
always funny

>>8665138
The guy in >>8665021 doesn't seem to be finding these issues. I'm prompting "oral, vaginal, anal" as he did and I'm getting everything but that.

>>8665143
I see that you're me. Idk how the fuck he's doing it because I can never get triple penetration to work but I'll have to try again tonight.

it might be model/lora dependent, but I get 3p no problem

>>8665145
>>8665143
Here's what I'm getting with the following prompt:
nekomata okayu, cat girl, female, strap-on, triple penetration, oral penetration, vaginal penetration, anal penetration, motion lines, motion blur, solo, masterpiece, best quality, very awa, absurdres, newest, detailed background, detailed eyes, detailed face, <lora:DetailedEyes_V3:1>
Negative prompt: worst quality, worst aesthetic, lowres, bad anatomy, bad hands, simple background, signature, logo, watermark, Patreon, Twitter, artist name, url, copyright, censored,

It's not 100% consistent, but it seems to be working more often than not. Maybe it's the NovaAnime model I'm using? https://civitai.com/models/376130/nova-anime-xl

>>8665149
Your pic is typically the thing I see. Even with spitroast it will begin by forgetting one of the boys or with gangbang it will duplicate the penis to fill the holes. It should be extremely simple, but its not. The model should be the culprit but 102d must be able to handle triple penetration surely.

>>8665149
It looks like a good starting point, even with the double strap-on that is most likely RNG.

>NovaAnime model
I will try this.

it's okay until you try to see where the bodies are

>>8665148
her age?

>>8665159
GigaToddlerfuck - A Toddlercon VN*
*All characters are 18+ even if they are depicted otherwise

>>8665148
I want to believe Goro, the Mortal Kombat guy, is below her.

oekaki so kino...

>>8665172
Following up with this, if anyone doesn't know oekaki and tegaki were Japanese art programs with certain limitations on things like resolution and color palettes which give them a distinct style. Also Danbooru has an art nouveau tag for some reason despite lacking tags for most other art movements?

Pic related is a quick test to show off the effects using the base NoobAI model.

* UNSUBSCRIBE

oreocookie :skull:

Since >>8665153 mentioned trying NovaAnime here's a quick rundown of Nova models.

Nova models are created by a model trainer named Crody: https://civitai.com/user/Crody/models . Each one is intended to have a strong in-built style, sacrificing flexibility for consistency. Because of that, there are a bunch of different models all with different built in styles.

NovaAnime is default anime one and the one that I typical use if I'm going to use a Nova model for anime gens.

Nova Anime3D is as the name implies trying to emulate a 3D anime style, as seen in newer anime and similar things.

Nova Cross is one of the newest ones and basically a mix of the above two for a 2.5D look.

Nova EXAnime is based on his SD1.5 Nova Hentai model. Apparently it was done completely separately than the Nova Anime model, with the training including a bunch of hentai, though it does seem like the final results are similar. I haven't experimented with this one much but might.

Nova Flat is a variation of Nova Anime focus on a flatter style while still having shading for colors.

Nova Orange is based on another set of models called OrangeMixes and is the opposite of Nova Flat, focused more on adding depth. I've seen this one used a lot in other merges on CivitAI.

Lastly, Nova Unreal is based on the model MeinaUnreal which was an SD1.5 model, and is a more animefied version of that model. A lot of the example images have a softer, more dream-like look to them.

Pic related is a test comparing all of them to WAI. Full res test at: https://files.catbox.moe/6bn0ff.png

:skull: ?

what's her endgame?

I think llmanon is cute. CUTE.

cant beat default style wai
its so good

>wai
lmoa

What are some good tags to make sex scenes feel more dynamic? I include "motion lines" and "motion blur" but is there anything more like that to make action more energetic?

>>8665206
>afterimage, emphasis lines, speed lines, bouncing breasts, ass ripple
Yep it's gooning time.

>>8665206
rough_sex as well

>>8665208
Good thinking anon! I had used emphasis lines for some things before but not all the time. I've definitely got to keep that and the others in mind though!

>>8665206
Still can't believe tentacle is not /h/

>>8665211
same but "normal" bestiality

>>8665211
I cannot find anything saying that but I apologize if it isn't.

>>8665209
You might be brilliant anon? It's hard to tell how much different "rough sex" versus just "sex" is making.

>>8665098
>>8665112
"fabric leotard" does something, but it's very artist dependent. I threw a bunch of random artists at it with a low step count to make it go faster, so it looks like ass but the result should be valid. Anything close to AIslop or "quality" really wants to have that shine, and some artists are also overfit on shiny clothes but in other cases it "works".

Here's another test, this time with multiple seeds. What's the consensus /hgg/ ? Does it look like "rough sex" is making the scene more intense?

Full res test: https://files.catbox.moe/emg00v.png

>>8665216
Not with this prompt anyway. Maybe if you didn't already add motion lines and ass ripple, or the fucking eye detail lora you might have seen a difference.

>>8665214
Ahah, thanks for doing this kind of testing! It was interesting to see your results. Good to know that sometimes the AI knows what "fabric" is at least. Which model is this?

>>8665216
The 'ass ripple' part seems to be a mistake

>>8665216
I think it's doing a lot of the things you have other prompt covering. I also think it helps with choking and hair pulling.

But it could also very well be snakeoil

>>8665222
102d custom - https://civitai.com/models/1201815?modelVersionId=1491533
it's about the closest you can get to base noob v-pred without requiring cfg rescale
seeing as every merge dilutes the model's knowledge somewhat (and adds shiny bias), it had a better chance of working

>>8665220
>>8665224
>>8665225
Yeah, all the other tags overwhelming the effect makes sense. Here's another one with everything removed except the sex tag, and comparing it with "happy sex" too thanks to >>8665225

Full res test: https://files.catbox.moe/zs45m3.png

>>8665228
Interesting! Thanks for the link, will check it out.

what's with all the shiny 2.5D spam? Some new attempt to kill /hgg/

hi

can you fucking stop spamming ugly eps models? you're hurting my eyes with grey sepia they're only capable of

its local

What's wrong with eps?

its grey

its sepia

>>8665177
>has an art noveau tag
>doesn't have a golden age of american illustration tag
having good taste is isolating

>>8665232
>Helpful anon does a helpful test
>Unhelpful anon unhelpfully seethes

>40:00
>41:00
I'm noticing

Speaking of helpful, one thing I got recently is the Infinite Image Browsing extension: https://github.com/zanllp/sd-webui-infinite-image-browsing

This lets you search images by metadata, which is super helpful if you have to find old gens. It's way easier to just type a character name or something else you remember from the prompt rather than going through all the old date folders trying to find what you're looking for.

Thank you, ChatGPT, you're being very helpful!

>>8665232
can't kill what's already dead

>finish 1 lora
>add 3 more to the list
help~

>>8665261
What are you training LORAs for? Artists, characters or something else?

>>8665263
that's so hi def i can smell her through my screen

>>8665263
I'm training a lora to surpass metal gear

>>8665251
Yeah this is by far the best extension for local. It's the reason why I'll never switch to comfy. You can also compare two images in the window if you drag and drop.

>>8665283
>You can also compare two images in the window if you drag and drop.
Wait what really?

>>8665263
Oh shit, I haven’t thought about Rabi in years!

>>8665290
Yes but only two at a time.

>>8665314
there is clearly 2 lies on that image

>>8665314
amazing WAI image

reforge got like 100 updates, wasn't it dead?

>>8665346
>is rael
WTF

>forgotten hero returns to save local
>https://github.com/Panchovix/stable-diffusion-webui-reForge/discussions/377
Any other such cases?

>>8665346
>>8665349

>>8665349
>>8665351
>still only supports SDXL
who the fuck cares? the future is chroma

>>8665357
>>>/g/

>>8665361
I'm sorry that your job stocking shelves at Reliance Retail doesn't let you buy a GPU with more than 8GB of VRAM

>>8665362
present your chroma gen, sir

>panchito returns
>already causing the axe wounds to piss and shit all over themselves
Missed that lil' nigga like you wouldn't believe

>>8665363
shut the fuck up retard

>>8665366
jej

>>8665366
kek

>>8665367
>>8665370
retard go the hell with SDXL too and stop replying

local will never improve as long as it's users are too poor to use models that isn't old busted SDXL with 4ch VAE CLIP TE and other obsolete tech cope we have to force people to upgrade hardware or we will be stuck on SDXL for ever

you heard it folks, the moment we all buy 3000$ gpus, best model will be created and provided to us instantly, by nvidia presumably, right?

What caused it THIS time?

>local out of nowhere

me, i am the cause

proofs?

/hdg/ will NEVER recover

>>8665382
it's called demand and offer you dumb socialist retard stop posting and embarrassing yourself

kys naishitter

uh oh useless overpriced gpu goy buyer meltie

nai is a pretty fair deal all things considered

can nai do video?

suuuuure

Wrong tab, bros.

wrong hole, bro

>>8665395
no

>>8665400
thought so

>>8665395
No, NAI cannot do 4 second loops of rotoscoped style animated characters awkwardly shuffling around with no object permanence and warping body parts and facial features.

keeeeek

>naifag cope

Did anyone ever remake old sd1.5/pony styles for Illustrious or Noob models?

I miss the Hades one, Nat the Lich, and Kotoyoshi Yumisuke... among others, but yeah.

Hell, I miss gayshit/makesomefuckingporn.

>>8665405
nope, no one ever did that

>>8665407
Well thats a shame, the Hades one lead to really striking shading. Really cool for flat colors/fully 2d images.

I can't stop genning BBC...

so true, saar!

I can't stop genning pregnant futas...

>>8665413
baaaaased

new fetish unlocked: bbc

what caused

That's it. Going back to /hdg/

What's the best model for the 2.5d AI-generated style? No it's not Wai. 102d 2.5d boost is good but I could have sworn there were many others. The one time I'm looking for it I can't find it.

>>8665443
1,5

>left /h/ for two weeks
>a brand new redditard lurking through civit slop and blogposting about it appeared
>Tried using a variety of materials I saw appear in some tags, but only some of them kind of worked sometimes.
>What's the best model for the 2.5d AI-generated style? No it's not Wai. 102d 2.5d boost is good but I could have sworn there were many others. The one time I'm looking for it I can't find it.
it's all so tiresome.

>>8665451
Fuck off retard. No one asked you to come back.

ah here it is here it is!

>>8665451
golden age of aesthetically pleasing images on /hdg/ is long gone. all genuine posters were banned by jannies. scatsloppers came instead. deal with it

Pancho will save us. The wheels are in motion. Trust.

>>8665455
Based. Every time anon asks what caused the decline, the answer is "jannies".

Oh I found it, it was coconut.
>>8665451
Why do you care what I do in the privacy of my own house? Why would this of all things tire you? Don't you get bored of policing this place?

>>8665457
>the answer is "jannies"
it's largely true althoughever
>>8665458
you're mistaking 4chan for your own house, fag

>>8665462
>ask for model recs to gen on my local AI box
>retard gets mad because someone asks a question
>decides to shit up the thread because someone somewhere is doing something he doesn't like
This isn't your house neither, retard. You're welcome to go back to where you came from, however.

>>8664451
What's your preferred CFG value? I only know that CFG++ should be 1-2 but I have no clue about DPM++ samplers.

localkeks are ridiculous lmao

>>8665465
It's the usual 5.

>>8665464
all probably true yet I still agree wit him
I think it's because of the AI slopstyle in everything he makes, sloppec checkpoint, way too many negs, too many quality tags, detail lora, etc

>>8665464
This is one of the most pathetic attempt at gaslighting/samefagging ever posted, please get better material.

>>8665472
You are just very retarded and can't see past your own headspace. Your stubby fingers could solve your own problem right now and alt F4 but you're too stupid to manage even that.

>>8665474
Did the janny rangeban your whole area and now you can't make any more slide threads, so you have to resort to this? That's pretty sad man.

>>8665476
Gonna cry?

>>8665494
nice areola
box?

>>8665206
Gotta change the composition
Flat shit is always boring, embrace weird angles

>>8665351
>guy lasted like a month "quitting"
jej many such cases
anyway he should sit down and unravel the mess

what mess

>want to remove a single tab in gradio
>it would require changing aprox 23 script files and 27 function aliases because of how messy the code is
>eventually resign to banning it in ublock

>>8665500
<https://files.catbox.moe/wz61hr.png>

>>8664384
meta plz

>>8665087
If you're that curious, here you go: <https://files.catbox.moe/c3cymq.png>

damn the oekaki didn't bake nicely
i think i need to reduce the dataset and retag it

>>8665547
how did the tes lora rebake go?

always a config issue

>>8665550
oh it was pretty good
i didn't keep it though kek

hm i wonder if resizing the images 3x nearest neighbor so they fit into proper buckets while still being pixelated would help

>>8665553
no oblivion flame atronach...

>>8665570
That one didn't want to work in the mix, but I do still have the datasets, I could bake it separately.

>>8665571
Oh, kinda figured it wouldn't but still unfortunate

>>8665560
that didn't do much
too bad

oh wait it did work spendidly
i just got that stupid fuckin glitch where loras of the same name aren't flushed out of cache and somehow get called even when deleted

>>8664749
catbox?

>a sexy woman dressed as a teacher in a classroom, a detailed sketch, ultra realistic, looking at viewer
good morning sirs

>looking at viewer
cuck pov...

>>8665604
you're the classroom hamster

Cuckolding/netorase is a very high IQ fetish — I don't expect the average person to "get it".

from side, looking at viewer, grin, middle finger, viewfinder

2boys, pov, chasitty cage, precum, 1girl, doggystyle, looking at viewer, middle finger, netorare, on bed, ahegao

/hgg/ doesn't stand for homo gay general tho

Is there an up to date Lora training guide for Illustrious/XL models? Wanna try baking one for an obscure character.
On that note, what's the bare minimum quality for an image to be worth including in the training data? I don't have all that much to work with.

the upscaling thing worked for eencya too
nice

>>8665662
that's a lot better, and all you did was upscale 3x?

>>8665666
nearest neighbor yeah
i use no upscale buckets so it just made them fit better at 1mp

>>8665669
if you're just doing nearest-neighbor you might as well let the training script upscale

>>8665678
it doesn't use nearest neighbor though
and the bucket distribution is shittier, it consolidates them way more

You know, I'll take the sepia v/e/rmins over unironic homos any day of the week

i bet you think about taking other men all day

>it helps with the lowres cgs too
bros... not like that... i don't wanna rebake 100 datasets...

>>8665687
then don't
doubt anyone cares about a 5% improvement, not to mention most of these artists already work well with no loras at all

lil bro i am not baking for people
i am baking for myself

>>8665692
based. sloppers will never get the joy of baking a 1% more faithful lora

>>8665689
why are you even here? just post your 20+ artist slop in hdg

Man from time to time this thread just goes to shit with no explanation
Guess I'll do the sensible thing and only ever come here when I have a question to ask

hello bwos hows it going

>>8665683
Are the homos in the thread with us?

>>8665703
Yes >>8665677

>>8665702
she looks like she had acne on her pussy and scratched it bloody, cool gen though

>>8665704
>1boy, 1girl
>homo

>>8665405
By Hades do you mean https://civitai.com/models/1653796/illustrious-from-hades ? It's a model based on IllustriousXL that seems to be continuing on the Hades name.

Pic related is a quick grid to demonstrate the model.

>>8665713
why is it so...

If he's not a ugly bastard or a muscle bound hulk with a big veiny girthy glistening circumcised penis, it's homo.

>>8665706
oops didn't notice that, my bad bwo

Is this guy pretending to be a bot or is it the real deal?

>>8665713
more likely meant a style lora based on the game's art
not some random shitmerge you can't even tell from hundreds of others

>>8665405
Looks like there's another Hades model that's specifically for Flat2D: https://civitai.com/models/1691607/illustrious-flat2d-from-hades

Again, pic related is a grid to show off the look and feel of it.

hades has some hot men...

>>8665720
there's your answer

>>8665405
If >>8665721 is correct and you just meant a LORA based on the game, there's like a ton of those. Here's a test of 3 of them:

https://civitai.com/models/1046386/hades-character-art-style-il-or-shrekman-game-styles

https://civitai.com/models/1417307/hades-game-style-and-characters-illustration-lora

https://civitai.com/models/1693974/hades-style-for-illustrious

For the one that's labeled FLUX, that's a misnomer. The CivitAI description says:

>ERRATA: I do realize the file name says FLUX, that was a mistake I made when setting up the training but the Lora IS MADE FOR IL so please don't mind the name and do let me know if it can be retroactively fixed.

yep, it's a bot

>>8665648
There are LORA training guides in the OP. Other than those you can search online and there are plenty. One I keep referring back to is https://civitai.com/articles/138/making-a-lora-is-like-baking-a-cake

>On that note, what's the bare minimum quality for an image to be worth including in the training data? I don't have all that much to work with.

Quality matters more than quantity for LORA training, so you want the best quality images you can get especially if there are only a few. You might want to carefully upscale what images you do have if they're low-res. You can also cut up an image if it's like a character profile and use each individual image to get a larger dataset. And you can double your dataset with flip augmentation too.

Pic related is an old table about LORA training with recommended amount of images and settings and stuff. It is old though so take it with a grain of salt.

>>8665560
How many buckets do people usually use when training LORAs? I've heard that having fewer buckets is better for LORA quality, but I always end up with a ton because my source images are always different resolutions and I'm not cropping or resizing them because I'm worried that would hurt more.

Do you decide on one size and make all your training images match that size, or do you choose like a handful of buckets and resize everything to those?

>anything v5
amazing guide lil llm bro

>>8665648
Quality how? Pixels? Just uscale.
Artistic quality? I baked on 2007-deviantart-ass images for one of my characters because there was nothing else and it just works.

>>8665511
Can't you change the gradio-based UIs like A1111/Forge/Reforge just by changing the ui-config.json in the root directory? Granted I've never tried to remove a tab like that so I'm not sure if they're there but if they are it might be as simple as flipping a "visible" flag from true to false.

>>8665346
>>8665349
>>8665351
>>8665348
Thanks for the heads up about this! I'm always hesitant to update in case something breaks though. Is anyone else in the thread using ReForge?

The discussion thread says the new update has 150+ commits and if it's performance improvements and such then that would be great, but everything in my install before the update is working at least.

>finally smartened up and got myself a script so i don't have to input every lora separately into prompt sr for testing

>>8665766
>I'm always hesitant to update in case something breaks
Same, which is why I just write up the commit I was using before and if something breaks I just revert it back and it usually works
Sometimes I have to do a fresh install in the previous commit though

lil bro learned about xyz plots

lil bro is a fucking retard that can't read

Lora training anons, what upscale/downscale option when training causes this shit? Maybe I fucked up the buckets.

>>8665783
bicubic downscale while bucketing did something like that for me

>BIcubic
uegh i hate this dei propaganda everywhere
what's next
homocubic

>prompt 4 girls in their section of the mask
>less than that appear
So tired of this.

good sunday to you my sars i bring to you more gfd

>>8665785
Total image files found: 111
Total size: 390.8 MB
Average file size: 3.5 MB
Largest file: 19.1 MB
Smallest file: 135.7 KB
Average dimensions: 2223x2427
Max dimensions: 5248x5248
Min dimensions: 968x792
Average megapixels: 6.06 MP
Max megapixels: 23.13 MP

Hmm, maybe having the highest res possible dataset is not ideal. Can I change the downscale method in EzScripts? (sd-scripts backend)

>>8665571
What config are you using for your character loras? The adamw batch size 1 vpred config?

>>8665801
By editing the code. Open sd-scripts dir in an IDE like VS Code and search for `cv2.INTER_CUBIC`. You'll find a few hits in the bucketing code.

>>8665802
yeh

>>8665662
So you ended up upscaling all those tiny pictures? I knew it wouldn't have worked as is.
>>8665805
Did you do anything special for that multicharacter tes lora too? Just keep tokens 1 at 1 repeat for each character?

>>8665811
I mean it worked well enough to use, it's just more accurate as is.
Yeah, nothing unusual. I just consolidated the character tags separately before putting them together. I don't keep tokens either.

DAILY MODEL WATCH

A couple interesting ones this time around.

First up is Ecliptica Etherea: https://civitai.com/models/1791900/ecliptica-etherea . This is an anime-influenced semi-realistic model. Seems fine if a more realistic 2.5D style is what you're going for.

Next, the most interesting one, IchigoMilk: https://civitai.com/models/1792053/ichigomilk . This is a cute model that seems geared towards a watercolor look, though to really see that influence you need to add "watercolor" to the prompt like the example images on CivitAI. Still gets the bigger, more rounded eyes cute look either way.

Next is Illustrious-Kinu-Blend: https://civitai.com/models/1792493/illustrious-kinu-blend . This is another one where a whole model has been used to emulate an artist style instead of a LORA, in this case the artist is Kinu Nishimura. The style will come across stronger with the trigger phrase "(Nishimura Kinu Style)" .

Lastly is UmetanaMix V2: https://civitai.com/models/1791455/umetanamix-v2 . This is another anime semi-realistic mix similar to the first one today. Seems to give similar results though this one leans more towards the anime look while Ecliptica Etherea leans more towards the realistic look.

Full res test: https://files.catbox.moe/4c2u3k.png

:skull: :skull: :skull:

>>8665799
The ear nibbling is very cute! And you've got a distinctive look here too. Good job anon!

>>8665815
>Ecliptica Etherea
Ok this might be the one I was looking for earlier thank you.

i fucking love /g/!!!!

>>8665815
No mention on how none of the models did the centaur right?

me neither

>>8665815
Following up with this to show off more of the style intended for models that want a specific style in the prompt, here's IchigoMilk with "watercolor" in the prompt like the example images on CivitAI.

>>8665822
You're welcome! Glad I could be helpful.

>>8665824
but check out those backgrounds and butiful default style!12!! you can just hit the generate botton and get all of that!

>>8665829
>>8665815
Likewise, for Kinu Blend here's how it looks with the recommended "(Nishimura Kinu Style)" in the prompt.

did he even upload a single /h/ pic in that week of botting?

>>8665803
>cv2.INTER_CUBIC
It's the function trim_and_resize_if_required() in train_utils.py right?
Wtf does the guy have a comment saying he wants to use "INTER_AREA" but still have INTER_CUBIC. Pure Insanity.

>>8665837
Yep. It was INTER_AREA originally, but changed to bicubic in the fork - for whatever reason.
Just change it back to INTER_AREA, that fixed the jagged lineart for me.

Cn tile continues to not work with loras
Someone lied to me

>>8665839
Alright, thanks for the help anon. Retraining the lora now, hopefully this will fix it.

>>8665855
I hope you removed all the old .npz files

>CN
ishyggydiggy

>>8665857
Yeah, I put a print line to make sure it was working correctly. Caught the latents and removed them, thanks for the heads up though.

>>8665859
It should work then. You can try to decode the newly created latents and check if they look fine.

>>8665853
Based a1111 bro who hasn't updated since 2023

a9fed bros, report

>>8665861
I'm using reforge man

> reforge
oh no no no
do we tell him?

>>8665871
Incorrect tranny, you should be using "nyo" to establish your dominance
Also can someone give me a non-facetious answer to this, please
What am I doing wrong?

>>8665873
I think the shitposters are appearing because you haven't posted a pic of your problem yet.

Any advice for better 69 scenes? I found that adding "cunnilingus" to the prompt does a better job of getting tongue to pussy than without it. Prompt for pic related is:

makoto nanaya, squirrel girl, female, (69, 69 position,:1.1) motion lines, motion blur, duo, ahegao, pussy juice, cunnilingus, nekomata okayu, cat girl, female, masterpiece, best quality, very awa, absurdres, newest, detailed background, detailed eyes, detailed face, <lora:DetailedEyes_V3:1>
Negative prompt: worst quality, worst aesthetic, lowres, bad anatomy, bad hands, simple background, signature, logo, watermark, Patreon, Twitter, artist name, url, copyright, censored,

><lora:DetailedEyes_V3:1>
bruh

>>8665880
you can remove pony from your PC, but you can't remove it from your gens

><lora:1GirlStanding_v5:1>

>>8665880
It's the XY schizo. You'd think someone with the power of comparison would not run a dozen negs by default.

In case the question is reposted, tags that require a specific facial expression (cunnilingus, ahegao) will inevitably conflict.
Add "tongue, saliva" and get rid of the ahegao

>>8665877
It's a classic "TypeError: 'NoneType' object is not iterable", so there's not really any pic to post
I just updated my reforge to the latest version, but it wasn't working in the previous one either so it has nothing to do with that
I have lora control (dynamic lora weights), which has had a similar problem before, but I've disabled it before I even started genning
I honestly have no idea what's up with this

>>8665884
Thank you kindly! Realized other things are better suited for /h/ but appreciate the advice here.

/h/ seems to be a very narrow content window. Does two girls on one guy fit or should that be somewhere else? Danbboru has a series of "cooperative" tags I'm trying to test out.

>>8665885
Does the console window have an error with more detail, like a stack trace?

>>8665885
Hmm idk but that looks like text to img tab. Have you tried using it in img2img?

>>8665886
There's a huuuuuge list, I can post it here if it's helpful
>>8665887
It's obviously i2i, as far as I'm aware it doesn't work on t2i right?

Anyone update reforge now that pancho is back on it, curious how bricked it's gonna be for a bit probably just make another setup for the new one so I don't fuck myself like every time I update.

>>8665885
That's the catchall for everything
Doesn't really help

>>8665889
Have you tried doing a simple test on a fresh reforge install?

>>8665889
>There's a huuuuuge list, I can post it here if it's helpful

If you go to the top level of the stack trace that should have what the initial exception was. You could dump the whole thing if it's fit in a post. Typically it's one exception and then a million things throw errors but if you just fix that one thing it'll fix it all.

>>8665885
>8gb vram
in your nvidia control panel is your cuda sysmem fallback policy set to prefer no fallback? I wonder if you're OOMing and it's giving you a non-specific error for whatever reason, but I don't use reforge so I dunno really

>>8665884
By the way trying this more and this advice was great! Thank you for being helpful anon.

>>8665897
Oh, good eyes! Yeah >>8665885 with that memory usage it might just be an OOM (out of memory) error or other memory issue.

Ya ever optimize your process and realize how time wasteful you were

>>8665892
Right, that's why I didn't post it
I'm technologically retarded so I don't even know what I should post
>>8665893
No, not yet
I might if I don't find any solutions
>>8665896
Picrel is the top level of the stack trace, it goes on and on and on from there lol
>>8665897
That's a good question I'll check it in a min

>>8665897
It's just set to driver default actually
Should I set it to something?

>>8665906
No, if it were set to "prefer no fallback" your gpu wouldn't use your system ram as back-up in case it ran out of vram, since it's set to default that's not the issue at least, or it shouldn't be

>>8665904
Yes. Which is why I will quit baking loras if this one doesn't work out.

>>8665910
Alright
I'll try to disable lora control to see if that's the issue

>>8665905
Considering it got to 4% before failing it seems like it might be running out of a resource like memory, if it was a misconfiguration or something I'd expect it to fail immediately rather than start for a bit and fail shortly after.

What's at the bottom of the message? It might have a more descriptive error message there.

>>8665902
Cheers! Also, if you can, try making 'cooperative cunnilingus' work, please? I'm trying on my end and it only gens one girl when it calls for two.

>it can look at images
Gemma...

>>8665923
I would but that would be /u/ , right? That's why I deleted my earlier post.

Also does anyone know a good way to remove the disembodied torso from pov images?

Cooperative cunny

>>8665927
>I would but that would be /u/, right?
True. Anyway, I made it work so no problem.

>disembodied torso
It's RNG.

>>8665929
Made that work too.

>>8665922
It was lora control apparently, disabling the extension solved the issue
Should have noticed that in the traceback it says:
"AttributeError: 'NoneType' object has no attribute 'timestep'
'NoneType' object has no attribute 'timestep' "

Is there any alternatives to lora control (basically something that allows me to schedule loras) that work on reforge?

>>8665936
Glad you figured it out anon! When you say LORA Control, which extension do you mean? The description at https://github.com/cheald/sd-webui-loractl says it was intended to replace https://github.com/a2569875/stable-diffusion-webui-composable-lora which it claims doesn't work in newer versions of the UI.

In any case:

>"AttributeError: 'NoneType' object has no attribute 'timestep'
>'NoneType' object has no attribute 'timestep' "

This sounds like there's an object being passed that's expected to have an attribute called "timestep" that's not there. From the name I'd guess that's the step number that the LORA is supposed be to applied or changed. Are you following an example or guide and do you have all the configurations and settings correct?

fucking lol

>>8665939
The original version of loractl doesn't seem to be compatible with reforge, so I'm using this one:
https://github.com/Panchovix/sd_webui_loractl_reforge_y

>>8665821
Thanks friend. Nice beerdragon.

>>8665942
What's your prompt like? It looks like the formatting for passing the parameters is a bit complex:

><lora:network_name[:te_multiplier[:unet_multiplier[:dyn_dim]]]>

If you aren't passing everything correctly then it could be causing an exception.

>>8665945
Thanks. And nice butt! I really like the expression you've been giving your catgirl too, it looks great. Are you using blush stickers instead of just blush for that look?

>>8665950
I wasn't using it at all in that one example, it was even off
Even then it gave me that issue, it only stopped once I removed the extension

>>8665950
i use the tag "slight blush circles" which seems to work a lot of the time.

>>8665927
Prompt male legs. Do a grid for this one.

>>8665936
>alternatives to schedule loras on reforge
Comfybro here. As a last-resort alternative to any kind of scheduling you can use the SDXL refiner flow, and just refine with the same model. Should let you change the lora/prompt/CN/whatever, and what happens under the hood is exactly the same as scheduling.

>>8665927
Depends a lot on the checkpoint, sometimes "disambodied penis" with "pov" works, otherwise you see the cutoff, or worse it adds another dude's dick to the scene. Sometimes negging "abs" and "pubic hair" works, if you can't see the girls' bottom halves.

Going by tags only it should be "pov crotch" in negs, but that never worked for me. Negs on illu/noob are pretty weak in general compared to positives. Maybe with negpip extension.

is that the famous videogen?

videogen be looking like those shitty mouth swap animations from 2006
now with 3000000% more rainforest destruction

>>8665944
Man I swear I get off more to /e/ posing than I do to sex these days. I thought porn was supposed to rot your brain? Nice gen.

>>8665990
Remember how bad and jank AI images were? Remember that was like, two years ago? Just wait, bro.

>>8665998
Remember how bad videogen was just a year ago? It wasn't even coherent.

>>8665990
>more rainforest destruction
I never understood this meme. Power generation outputs CO2, right? That's plant food. You're feeding the rainforest when you generate.

>>8665998
But AI images are still bad and jank?

>>8666002
You must've been top of your fuckin class.

Turns out "cooperative grinding" doesn't work on a third pussy, and "cooperative tribadism/scissoring" isn't a thing. Who knew?

>>8666005
He clearly was since you couldn't even refute him. Haven't seen someone btfo'd this hard in a while.

>>8666009
Who's speaking here? Is somebody speaking?

>>8666010
>literally crying
lol
lmao

>>8666003
Only the ones made with novelai v4.5

>>8666011
Your sister's cunt!

>>8665976
>refiner
Whenever I change models it takes like 3 minutes because I'm poor and don't have an ssd
Using refiner with the same model is actually a very interesting idea, gotta give it some runs

>>8665990
It's a fun tech but local videogen still looks like fucking poop and I'm genuinely surprised to see so many people liking it, even on /a/. Some shit like google veo looks actually impressive, especially combined with sound.

>>8666018
>especially combined with sound
Every video I've seen made with veo makes the people sound as if they were underwater.
Still, something like that but for anime (and that allowed NSFW content) would be great.

>>8666018
At the same time this was acceptable once here
My own hypernetwork time stock kek

>>8666021
Not like any of us would be able to run it.

>>8665719
where do i subscribe?

white man was here
how do you know?
cuckold porn

>>8665422
BIBICIIIIII

>>8666035
why is blud looking straight out Clash of Clans :skull:

>>8666037
I hate a substantial portion of whites due to that

indian man was here
how do you know?
HELLO SAAR

>>8666044
I hate all indians due to that
Also they're big into cuckold shit too

i would say 95% of cuck shit is white pypol
same with chastity and shit

>>8666046
did you hop off the /pol/ train or sumtin' mister? we loooooove black dick here.

I thought you n-words were back to /hdg/ already

>>8666048
Never even went to pol
Go back to hdg

>>8666052
hmmmm, nyo :3

>>8665997
Same
I'll be genning a cute girl looking a little lewd, getting rock hard and then I put her in actual sexual situations and I just get bored
I think with AI you need to show some restraint to retain interest, if you just go all out degeneracy you kind of become disillusioned, where as with /e/ material you still get to use your imagination

>nyo
lol
lmao even

>>8665815
stop comparing slop with slop
at least use Noob 1.0 instead of wai of all things

bbc jumpscare

/hdg/ wonned

>>8665997
sex was always fucking boring, tease is where it's at
sadly ai can barely do it really on point, especially when it comes to expressions

Why do niggers and faggots always walk side by side

>>8666056
saar noob 1.0 doens't do butiful sexy anime realism with detailed background

>>8666059
>>8666054
For me it's the implications of sex rather than sex itself, tally marks somewhere, used condoms under clothes, wet clothes around the crotch, etc

>be looking at gelbooru from the ass end, page 3141 or whatever
>intensely more kino than all the new shit i've been looking at
When did it go wrong?

>>8666057
oh my science!

>>8666054
>use your imagination
Yeah this might be it. I enjoy out of frame censoring for this reason also.
>>8666059
But like anon said, this requires more effort on your part. AI thinks you want sex but you need to be clever with your prompting to get good teases.

>>8666064
>wet clothes around the crotch
What's the prompt for this?

>>8666065
fotm gacha fanarts were a mistake

>>8666068
"pussy juice stain" will put you almost there

>itt: anons rediscovering why fine art consists mostly of tasteful nudes

>>8666071
I just remember getting cucked because the "stain" instead turned into blood stain and I never tried it again.

>>8666065
smaller artists are unironically better than the big ones
the big ones on danbooru just mass produce the same gunk with rigid compositions and boring scenarios, they might as well be AI sloppers

>>8666074
Agreed
On this topic, what small artists do you guys like?

>>8666072
didn't know my gens were subject to moral and church censorship but sure ig

>>8666073
"blood" in negs?

>>8666073
It worked for me because I also used related tags for it, perhaps you needed to give the slopmachine more "context" with those

>>8666074
It's not even big v small or anything, it's just old v new. Even guys who used to make kino back in dem day produce gachaslop that looks the same as the other 1000000 gachaslop artists. It's that 2.2d.

>>8666076
It never happened, but should have, and your should too

>>8666075
I like void 0, for some kino lighting

>>8666075
>what small artists do you guys like?
small like there is just few images of his style or small like no one seems to know him?

>>8666091
small as in having a micropenis

>>8666088
Ayo that's awesome, I'll add him to my list
>>8666091
Both
Less than 200 images on danbooru + being mostly unknown

>>8666088
Woah good stuff anon.